Upload 9 files

benchmarks.py

@@ -0,0 +1,142 @@
+import argparse
+import platform
+import sys
+import time
+from pathlib import Path
+
+import pandas as pd
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+# ROOT = ROOT.relative_to(Path.cwd())  # relative
+
+import export
+from models.experimental import attempt_load
+from models.yolo import SegmentationModel
+from segment.val import run as val_seg
+from utils import notebook_init
+from utils.general import LOGGER, check_yaml, file_size, print_args
+from utils.torch_utils import select_device
+from val import run as val_det
+
+
+def run(
+        weights=ROOT / 'yolo.pt',  # weights path
+        imgsz=640,  # inference size (pixels)
+        batch_size=1,  # batch size
+        data=ROOT / 'data/coco.yaml',  # dataset.yaml path
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        half=False,  # use FP16 half-precision inference
+        test=False,  # test exports only
+        pt_only=False,  # test PyTorch only
+        hard_fail=False,  # throw error on benchmark failure
+):
+    y, t = [], time.time()
+    device = select_device(device)
+    model_type = type(attempt_load(weights, fuse=False))  # DetectionModel, SegmentationModel, etc.
+    for i, (name, f, suffix, cpu, gpu) in export.export_formats().iterrows():  # index, (name, file, suffix, CPU, GPU)
+        try:
+            assert i not in (9, 10), 'inference not supported'  # Edge TPU and TF.js are unsupported
+            assert i != 5 or platform.system() == 'Darwin', 'inference only supported on macOS>=10.13'  # CoreML
+            if 'cpu' in device.type:
+                assert cpu, 'inference not supported on CPU'
+            if 'cuda' in device.type:
+                assert gpu, 'inference not supported on GPU'
+
+            # Export
+            if f == '-':
+                w = weights  # PyTorch format
+            else:
+                w = export.run(weights=weights, imgsz=[imgsz], include=[f], device=device, half=half)[-1]  # all others
+            assert suffix in str(w), 'export failed'
+
+            # Validate
+            if model_type == SegmentationModel:
+                result = val_seg(data, w, batch_size, imgsz, plots=False, device=device, task='speed', half=half)
+                metric = result[0][7]  # (box(p, r, map50, map), mask(p, r, map50, map), *loss(box, obj, cls))
+            else:  # DetectionModel:
+                result = val_det(data, w, batch_size, imgsz, plots=False, device=device, task='speed', half=half)
+                metric = result[0][3]  # (p, r, map50, map, *loss(box, obj, cls))
+            speed = result[2][1]  # times (preprocess, inference, postprocess)
+            y.append([name, round(file_size(w), 1), round(metric, 4), round(speed, 2)])  # MB, mAP, t_inference
+        except Exception as e:
+            if hard_fail:
+                assert type(e) is AssertionError, f'Benchmark --hard-fail for {name}: {e}'
+            LOGGER.warning(f'WARNING ⚠️ Benchmark failure for {name}: {e}')
+            y.append([name, None, None, None])  # mAP, t_inference
+        if pt_only and i == 0:
+            break  # break after PyTorch
+
+    # Print results
+    LOGGER.info('\n')
+    parse_opt()
+    notebook_init()  # print system info
+    c = ['Format', 'Size (MB)', 'mAP50-95', 'Inference time (ms)'] if map else ['Format', 'Export', '', '']
+    py = pd.DataFrame(y, columns=c)
+    LOGGER.info(f'\nBenchmarks complete ({time.time() - t:.2f}s)')
+    LOGGER.info(str(py if map else py.iloc[:, :2]))
+    if hard_fail and isinstance(hard_fail, str):
+        metrics = py['mAP50-95'].array  # values to compare to floor
+        floor = eval(hard_fail)  # minimum metric floor to pass
+        assert all(x > floor for x in metrics if pd.notna(x)), f'HARD FAIL: mAP50-95 < floor {floor}'
+    return py
+
+
+def test(
+        weights=ROOT / 'yolo.pt',  # weights path
+        imgsz=640,  # inference size (pixels)
+        batch_size=1,  # batch size
+        data=ROOT / 'data/coco128.yaml',  # dataset.yaml path
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        half=False,  # use FP16 half-precision inference
+        test=False,  # test exports only
+        pt_only=False,  # test PyTorch only
+        hard_fail=False,  # throw error on benchmark failure
+):
+    y, t = [], time.time()
+    device = select_device(device)
+    for i, (name, f, suffix, gpu) in export.export_formats().iterrows():  # index, (name, file, suffix, gpu-capable)
+        try:
+            w = weights if f == '-' else \
+                export.run(weights=weights, imgsz=[imgsz], include=[f], device=device, half=half)[-1]  # weights
+            assert suffix in str(w), 'export failed'
+            y.append([name, True])
+        except Exception:
+            y.append([name, False])  # mAP, t_inference
+
+    # Print results
+    LOGGER.info('\n')
+    parse_opt()
+    notebook_init()  # print system info
+    py = pd.DataFrame(y, columns=['Format', 'Export'])
+    LOGGER.info(f'\nExports complete ({time.time() - t:.2f}s)')
+    LOGGER.info(str(py))
+    return py
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--weights', type=str, default=ROOT / 'yolo.pt', help='weights path')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--test', action='store_true', help='test exports only')
+    parser.add_argument('--pt-only', action='store_true', help='test PyTorch only')
+    parser.add_argument('--hard-fail', nargs='?', const=True, default=False, help='Exception on error or < min metric')
+    opt = parser.parse_args()
+    opt.data = check_yaml(opt.data)  # check YAML
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    test(**vars(opt)) if opt.test else run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

export.py

@@ -0,0 +1,606 @@
+import argparse
+import contextlib
+import json
+import os
+import platform
+import re
+import subprocess
+import sys
+import time
+import warnings
+from pathlib import Path
+
+import pandas as pd
+import torch
+from torch.utils.mobile_optimizer import optimize_for_mobile
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+if platform.system() != 'Windows':
+    ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.experimental import attempt_load
+from models.yolo import ClassificationModel, Detect, DetectionModel, SegmentationModel
+from utils.dataloaders import LoadImages
+from utils.general import (LOGGER, Profile, check_dataset, check_img_size, check_requirements, check_version,
+                           check_yaml, colorstr, file_size, get_default_args, print_args, url2file, yaml_save)
+from utils.torch_utils import select_device, smart_inference_mode
+
+MACOS = platform.system() == 'Darwin'  # macOS environment
+
+
+def export_formats():
+    # YOLO export formats
+    x = [
+        ['PyTorch', '-', '.pt', True, True],
+        ['TorchScript', 'torchscript', '.torchscript', True, True],
+        ['ONNX', 'onnx', '.onnx', True, True],
+        ['OpenVINO', 'openvino', '_openvino_model', True, False],
+        ['TensorRT', 'engine', '.engine', False, True],
+        ['CoreML', 'coreml', '.mlmodel', True, False],
+        ['TensorFlow SavedModel', 'saved_model', '_saved_model', True, True],
+        ['TensorFlow GraphDef', 'pb', '.pb', True, True],
+        ['TensorFlow Lite', 'tflite', '.tflite', True, False],
+        ['TensorFlow Edge TPU', 'edgetpu', '_edgetpu.tflite', False, False],
+        ['TensorFlow.js', 'tfjs', '_web_model', False, False],
+        ['PaddlePaddle', 'paddle', '_paddle_model', True, True],]
+    return pd.DataFrame(x, columns=['Format', 'Argument', 'Suffix', 'CPU', 'GPU'])
+
+
+def try_export(inner_func):
+    # YOLO export decorator, i..e @try_export
+    inner_args = get_default_args(inner_func)
+
+    def outer_func(*args, **kwargs):
+        prefix = inner_args['prefix']
+        try:
+            with Profile() as dt:
+                f, model = inner_func(*args, **kwargs)
+            LOGGER.info(f'{prefix} export success ✅ {dt.t:.1f}s, saved as {f} ({file_size(f):.1f} MB)')
+            return f, model
+        except Exception as e:
+            LOGGER.info(f'{prefix} export failure ❌ {dt.t:.1f}s: {e}')
+            return None, None
+
+    return outer_func
+
+
+@try_export
+def export_torchscript(model, im, file, optimize, prefix=colorstr('TorchScript:')):
+    # YOLO TorchScript model export
+    LOGGER.info(f'\n{prefix} starting export with torch {torch.__version__}...')
+    f = file.with_suffix('.torchscript')
+
+    ts = torch.jit.trace(model, im, strict=False)
+    d = {"shape": im.shape, "stride": int(max(model.stride)), "names": model.names}
+    extra_files = {'config.txt': json.dumps(d)}  # torch._C.ExtraFilesMap()
+    if optimize:  # https://pytorch.org/tutorials/recipes/mobile_interpreter.html
+        optimize_for_mobile(ts)._save_for_lite_interpreter(str(f), _extra_files=extra_files)
+    else:
+        ts.save(str(f), _extra_files=extra_files)
+    return f, None
+
+
+@try_export
+def export_onnx(model, im, file, opset, dynamic, simplify, prefix=colorstr('ONNX:')):
+    # YOLO ONNX export
+    check_requirements('onnx')
+    import onnx
+
+    LOGGER.info(f'\n{prefix} starting export with onnx {onnx.__version__}...')
+    f = file.with_suffix('.onnx')
+
+    output_names = ['output0', 'output1'] if isinstance(model, SegmentationModel) else ['output0']
+    if dynamic:
+        dynamic = {'images': {0: 'batch', 2: 'height', 3: 'width'}}  # shape(1,3,640,640)
+        if isinstance(model, SegmentationModel):
+            dynamic['output0'] = {0: 'batch', 1: 'anchors'}  # shape(1,25200,85)
+            dynamic['output1'] = {0: 'batch', 2: 'mask_height', 3: 'mask_width'}  # shape(1,32,160,160)
+        elif isinstance(model, DetectionModel):
+            dynamic['output0'] = {0: 'batch', 1: 'anchors'}  # shape(1,25200,85)
+
+    torch.onnx.export(
+        model.cpu() if dynamic else model,  # --dynamic only compatible with cpu
+        im.cpu() if dynamic else im,
+        f,
+        verbose=False,
+        opset_version=opset,
+        do_constant_folding=True,
+        input_names=['images'],
+        output_names=output_names,
+        dynamic_axes=dynamic or None)
+
+    # Checks
+    model_onnx = onnx.load(f)  # load onnx model
+    onnx.checker.check_model(model_onnx)  # check onnx model
+
+    # Metadata
+    d = {'stride': int(max(model.stride)), 'names': model.names}
+    for k, v in d.items():
+        meta = model_onnx.metadata_props.add()
+        meta.key, meta.value = k, str(v)
+    onnx.save(model_onnx, f)
+
+    # Simplify
+    if simplify:
+        try:
+            cuda = torch.cuda.is_available()
+            check_requirements(('onnxruntime-gpu' if cuda else 'onnxruntime', 'onnx-simplifier>=0.4.1'))
+            import onnxsim
+
+            LOGGER.info(f'{prefix} simplifying with onnx-simplifier {onnxsim.__version__}...')
+            model_onnx, check = onnxsim.simplify(model_onnx)
+            assert check, 'assert check failed'
+            onnx.save(model_onnx, f)
+        except Exception as e:
+            LOGGER.info(f'{prefix} simplifier failure: {e}')
+    return f, model_onnx
+
+
+@try_export
+def export_openvino(file, metadata, half, prefix=colorstr('OpenVINO:')):
+    # YOLO OpenVINO export
+    check_requirements('openvino-dev')  # requires openvino-dev: https://pypi.org/project/openvino-dev/
+    import openvino.inference_engine as ie
+
+    LOGGER.info(f'\n{prefix} starting export with openvino {ie.__version__}...')
+    f = str(file).replace('.pt', f'_openvino_model{os.sep}')
+
+    cmd = f"mo --input_model {file.with_suffix('.onnx')} --output_dir {f} --data_type {'FP16' if half else 'FP32'}"
+    subprocess.run(cmd.split(), check=True, env=os.environ)  # export
+    yaml_save(Path(f) / file.with_suffix('.yaml').name, metadata)  # add metadata.yaml
+    return f, None
+
+
+@try_export
+def export_paddle(model, im, file, metadata, prefix=colorstr('PaddlePaddle:')):
+    # YOLO Paddle export
+    check_requirements(('paddlepaddle', 'x2paddle'))
+    import x2paddle
+    from x2paddle.convert import pytorch2paddle
+
+    LOGGER.info(f'\n{prefix} starting export with X2Paddle {x2paddle.__version__}...')
+    f = str(file).replace('.pt', f'_paddle_model{os.sep}')
+
+    pytorch2paddle(module=model, save_dir=f, jit_type='trace', input_examples=[im])  # export
+    yaml_save(Path(f) / file.with_suffix('.yaml').name, metadata)  # add metadata.yaml
+    return f, None
+
+
+@try_export
+def export_coreml(model, im, file, int8, half, prefix=colorstr('CoreML:')):
+    # YOLO CoreML export
+    check_requirements('coremltools')
+    import coremltools as ct
+
+    LOGGER.info(f'\n{prefix} starting export with coremltools {ct.__version__}...')
+    f = file.with_suffix('.mlmodel')
+
+    ts = torch.jit.trace(model, im, strict=False)  # TorchScript model
+    ct_model = ct.convert(ts, inputs=[ct.ImageType('image', shape=im.shape, scale=1 / 255, bias=[0, 0, 0])])
+    bits, mode = (8, 'kmeans_lut') if int8 else (16, 'linear') if half else (32, None)
+    if bits < 32:
+        if MACOS:  # quantization only supported on macOS
+            with warnings.catch_warnings():
+                warnings.filterwarnings("ignore", category=DeprecationWarning)  # suppress numpy==1.20 float warning
+                ct_model = ct.models.neural_network.quantization_utils.quantize_weights(ct_model, bits, mode)
+        else:
+            print(f'{prefix} quantization only supported on macOS, skipping...')
+    ct_model.save(f)
+    return f, ct_model
+
+
+@try_export
+def export_engine(model, im, file, half, dynamic, simplify, workspace=4, verbose=False, prefix=colorstr('TensorRT:')):
+    # YOLO TensorRT export https://developer.nvidia.com/tensorrt
+    assert im.device.type != 'cpu', 'export running on CPU but must be on GPU, i.e. `python export.py --device 0`'
+    try:
+        import tensorrt as trt
+    except Exception:
+        if platform.system() == 'Linux':
+            check_requirements('nvidia-tensorrt', cmds='-U --index-url https://pypi.ngc.nvidia.com')
+        import tensorrt as trt
+
+    if trt.__version__[0] == '7':  # TensorRT 7 handling https://github.com/ultralytics/yolov5/issues/6012
+        grid = model.model[-1].anchor_grid
+        model.model[-1].anchor_grid = [a[..., :1, :1, :] for a in grid]
+        export_onnx(model, im, file, 12, dynamic, simplify)  # opset 12
+        model.model[-1].anchor_grid = grid
+    else:  # TensorRT >= 8
+        check_version(trt.__version__, '8.0.0', hard=True)  # require tensorrt>=8.0.0
+        export_onnx(model, im, file, 12, dynamic, simplify)  # opset 12
+    onnx = file.with_suffix('.onnx')
+
+    LOGGER.info(f'\n{prefix} starting export with TensorRT {trt.__version__}...')
+    assert onnx.exists(), f'failed to export ONNX file: {onnx}'
+    f = file.with_suffix('.engine')  # TensorRT engine file
+    logger = trt.Logger(trt.Logger.INFO)
+    if verbose:
+        logger.min_severity = trt.Logger.Severity.VERBOSE
+
+    builder = trt.Builder(logger)
+    config = builder.create_builder_config()
+    config.max_workspace_size = workspace * 1 << 30
+    # config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, workspace << 30)  # fix TRT 8.4 deprecation notice
+
+    flag = (1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
+    network = builder.create_network(flag)
+    parser = trt.OnnxParser(network, logger)
+    if not parser.parse_from_file(str(onnx)):
+        raise RuntimeError(f'failed to load ONNX file: {onnx}')
+
+    inputs = [network.get_input(i) for i in range(network.num_inputs)]
+    outputs = [network.get_output(i) for i in range(network.num_outputs)]
+    for inp in inputs:
+        LOGGER.info(f'{prefix} input "{inp.name}" with shape{inp.shape} {inp.dtype}')
+    for out in outputs:
+        LOGGER.info(f'{prefix} output "{out.name}" with shape{out.shape} {out.dtype}')
+
+    if dynamic:
+        if im.shape[0] <= 1:
+            LOGGER.warning(f"{prefix} WARNING ⚠️ --dynamic model requires maximum --batch-size argument")
+        profile = builder.create_optimization_profile()
+        for inp in inputs:
+            profile.set_shape(inp.name, (1, *im.shape[1:]), (max(1, im.shape[0] // 2), *im.shape[1:]), im.shape)
+        config.add_optimization_profile(profile)
+
+    LOGGER.info(f'{prefix} building FP{16 if builder.platform_has_fast_fp16 and half else 32} engine as {f}')
+    if builder.platform_has_fast_fp16 and half:
+        config.set_flag(trt.BuilderFlag.FP16)
+    with builder.build_engine(network, config) as engine, open(f, 'wb') as t:
+        t.write(engine.serialize())
+    return f, None
+
+
+@try_export
+def export_saved_model(model,
+                       im,
+                       file,
+                       dynamic,
+                       tf_nms=False,
+                       agnostic_nms=False,
+                       topk_per_class=100,
+                       topk_all=100,
+                       iou_thres=0.45,
+                       conf_thres=0.25,
+                       keras=False,
+                       prefix=colorstr('TensorFlow SavedModel:')):
+    # YOLO TensorFlow SavedModel export
+    try:
+        import tensorflow as tf
+    except Exception:
+        check_requirements(f"tensorflow{'' if torch.cuda.is_available() else '-macos' if MACOS else '-cpu'}")
+        import tensorflow as tf
+    from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
+
+    from models.tf import TFModel
+
+    LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...')
+    f = str(file).replace('.pt', '_saved_model')
+    batch_size, ch, *imgsz = list(im.shape)  # BCHW
+
+    tf_model = TFModel(cfg=model.yaml, model=model, nc=model.nc, imgsz=imgsz)
+    im = tf.zeros((batch_size, *imgsz, ch))  # BHWC order for TensorFlow
+    _ = tf_model.predict(im, tf_nms, agnostic_nms, topk_per_class, topk_all, iou_thres, conf_thres)
+    inputs = tf.keras.Input(shape=(*imgsz, ch), batch_size=None if dynamic else batch_size)
+    outputs = tf_model.predict(inputs, tf_nms, agnostic_nms, topk_per_class, topk_all, iou_thres, conf_thres)
+    keras_model = tf.keras.Model(inputs=inputs, outputs=outputs)
+    keras_model.trainable = False
+    keras_model.summary()
+    if keras:
+        keras_model.save(f, save_format='tf')
+    else:
+        spec = tf.TensorSpec(keras_model.inputs[0].shape, keras_model.inputs[0].dtype)
+        m = tf.function(lambda x: keras_model(x))  # full model
+        m = m.get_concrete_function(spec)
+        frozen_func = convert_variables_to_constants_v2(m)
+        tfm = tf.Module()
+        tfm.__call__ = tf.function(lambda x: frozen_func(x)[:4] if tf_nms else frozen_func(x), [spec])
+        tfm.__call__(im)
+        tf.saved_model.save(tfm,
+                            f,
+                            options=tf.saved_model.SaveOptions(experimental_custom_gradients=False) if check_version(
+                                tf.__version__, '2.6') else tf.saved_model.SaveOptions())
+    return f, keras_model
+
+
+@try_export
+def export_pb(keras_model, file, prefix=colorstr('TensorFlow GraphDef:')):
+    # YOLO TensorFlow GraphDef *.pb export https://github.com/leimao/Frozen_Graph_TensorFlow
+    import tensorflow as tf
+    from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
+
+    LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...')
+    f = file.with_suffix('.pb')
+
+    m = tf.function(lambda x: keras_model(x))  # full model
+    m = m.get_concrete_function(tf.TensorSpec(keras_model.inputs[0].shape, keras_model.inputs[0].dtype))
+    frozen_func = convert_variables_to_constants_v2(m)
+    frozen_func.graph.as_graph_def()
+    tf.io.write_graph(graph_or_graph_def=frozen_func.graph, logdir=str(f.parent), name=f.name, as_text=False)
+    return f, None
+
+
+@try_export
+def export_tflite(keras_model, im, file, int8, data, nms, agnostic_nms, prefix=colorstr('TensorFlow Lite:')):
+    # YOLOv5 TensorFlow Lite export
+    import tensorflow as tf
+
+    LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...')
+    batch_size, ch, *imgsz = list(im.shape)  # BCHW
+    f = str(file).replace('.pt', '-fp16.tflite')
+
+    converter = tf.lite.TFLiteConverter.from_keras_model(keras_model)
+    converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS]
+    converter.target_spec.supported_types = [tf.float16]
+    converter.optimizations = [tf.lite.Optimize.DEFAULT]
+    if int8:
+        from models.tf import representative_dataset_gen
+        dataset = LoadImages(check_dataset(check_yaml(data))['train'], img_size=imgsz, auto=False)
+        converter.representative_dataset = lambda: representative_dataset_gen(dataset, ncalib=100)
+        converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+        converter.target_spec.supported_types = []
+        converter.inference_input_type = tf.uint8  # or tf.int8
+        converter.inference_output_type = tf.uint8  # or tf.int8
+        converter.experimental_new_quantizer = True
+        f = str(file).replace('.pt', '-int8.tflite')
+    if nms or agnostic_nms:
+        converter.target_spec.supported_ops.append(tf.lite.OpsSet.SELECT_TF_OPS)
+
+    tflite_model = converter.convert()
+    open(f, "wb").write(tflite_model)
+    return f, None
+
+
+@try_export
+def export_edgetpu(file, prefix=colorstr('Edge TPU:')):
+    # YOLO Edge TPU export https://coral.ai/docs/edgetpu/models-intro/
+    cmd = 'edgetpu_compiler --version'
+    help_url = 'https://coral.ai/docs/edgetpu/compiler/'
+    assert platform.system() == 'Linux', f'export only supported on Linux. See {help_url}'
+    if subprocess.run(f'{cmd} >/dev/null', shell=True).returncode != 0:
+        LOGGER.info(f'\n{prefix} export requires Edge TPU compiler. Attempting install from {help_url}')
+        sudo = subprocess.run('sudo --version >/dev/null', shell=True).returncode == 0  # sudo installed on system
+        for c in (
+                'curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -',
+                'echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list',
+                'sudo apt-get update', 'sudo apt-get install edgetpu-compiler'):
+            subprocess.run(c if sudo else c.replace('sudo ', ''), shell=True, check=True)
+    ver = subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().split()[-1]
+
+    LOGGER.info(f'\n{prefix} starting export with Edge TPU compiler {ver}...')
+    f = str(file).replace('.pt', '-int8_edgetpu.tflite')  # Edge TPU model
+    f_tfl = str(file).replace('.pt', '-int8.tflite')  # TFLite model
+
+    cmd = f"edgetpu_compiler -s -d -k 10 --out_dir {file.parent} {f_tfl}"
+    subprocess.run(cmd.split(), check=True)
+    return f, None
+
+
+@try_export
+def export_tfjs(file, prefix=colorstr('TensorFlow.js:')):
+    # YOLO TensorFlow.js export
+    check_requirements('tensorflowjs')
+    import tensorflowjs as tfjs
+
+    LOGGER.info(f'\n{prefix} starting export with tensorflowjs {tfjs.__version__}...')
+    f = str(file).replace('.pt', '_web_model')  # js dir
+    f_pb = file.with_suffix('.pb')  # *.pb path
+    f_json = f'{f}/model.json'  # *.json path
+
+    cmd = f'tensorflowjs_converter --input_format=tf_frozen_model ' \
+          f'--output_node_names=Identity,Identity_1,Identity_2,Identity_3 {f_pb} {f}'
+    subprocess.run(cmd.split())
+
+    json = Path(f_json).read_text()
+    with open(f_json, 'w') as j:  # sort JSON Identity_* in ascending order
+        subst = re.sub(
+            r'{"outputs": {"Identity.?.?": {"name": "Identity.?.?"}, '
+            r'"Identity.?.?": {"name": "Identity.?.?"}, '
+            r'"Identity.?.?": {"name": "Identity.?.?"}, '
+            r'"Identity.?.?": {"name": "Identity.?.?"}}}', r'{"outputs": {"Identity": {"name": "Identity"}, '
+            r'"Identity_1": {"name": "Identity_1"}, '
+            r'"Identity_2": {"name": "Identity_2"}, '
+            r'"Identity_3": {"name": "Identity_3"}}}', json)
+        j.write(subst)
+    return f, None
+
+
+def add_tflite_metadata(file, metadata, num_outputs):
+    # Add metadata to *.tflite models per https://www.tensorflow.org/lite/models/convert/metadata
+    with contextlib.suppress(ImportError):
+        # check_requirements('tflite_support')
+        from tflite_support import flatbuffers
+        from tflite_support import metadata as _metadata
+        from tflite_support import metadata_schema_py_generated as _metadata_fb
+
+        tmp_file = Path('/tmp/meta.txt')
+        with open(tmp_file, 'w') as meta_f:
+            meta_f.write(str(metadata))
+
+        model_meta = _metadata_fb.ModelMetadataT()
+        label_file = _metadata_fb.AssociatedFileT()
+        label_file.name = tmp_file.name
+        model_meta.associatedFiles = [label_file]
+
+        subgraph = _metadata_fb.SubGraphMetadataT()
+        subgraph.inputTensorMetadata = [_metadata_fb.TensorMetadataT()]
+        subgraph.outputTensorMetadata = [_metadata_fb.TensorMetadataT()] * num_outputs
+        model_meta.subgraphMetadata = [subgraph]
+
+        b = flatbuffers.Builder(0)
+        b.Finish(model_meta.Pack(b), _metadata.MetadataPopulator.METADATA_FILE_IDENTIFIER)
+        metadata_buf = b.Output()
+
+        populator = _metadata.MetadataPopulator.with_model_file(file)
+        populator.load_metadata_buffer(metadata_buf)
+        populator.load_associated_files([str(tmp_file)])
+        populator.populate()
+        tmp_file.unlink()
+
+
+@smart_inference_mode()
+def run(
+        data=ROOT / 'data/coco.yaml',  # 'dataset.yaml path'
+        weights=ROOT / 'yolo.pt',  # weights path
+        imgsz=(640, 640),  # image (height, width)
+        batch_size=1,  # batch size
+        device='cpu',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        include=('torchscript', 'onnx'),  # include formats
+        half=False,  # FP16 half-precision export
+        inplace=False,  # set YOLO Detect() inplace=True
+        keras=False,  # use Keras
+        optimize=False,  # TorchScript: optimize for mobile
+        int8=False,  # CoreML/TF INT8 quantization
+        dynamic=False,  # ONNX/TF/TensorRT: dynamic axes
+        simplify=False,  # ONNX: simplify model
+        opset=12,  # ONNX: opset version
+        verbose=False,  # TensorRT: verbose log
+        workspace=4,  # TensorRT: workspace size (GB)
+        nms=False,  # TF: add NMS to model
+        agnostic_nms=False,  # TF: add agnostic NMS to model
+        topk_per_class=100,  # TF.js NMS: topk per class to keep
+        topk_all=100,  # TF.js NMS: topk for all classes to keep
+        iou_thres=0.45,  # TF.js NMS: IoU threshold
+        conf_thres=0.25,  # TF.js NMS: confidence threshold
+):
+    t = time.time()
+    include = [x.lower() for x in include]  # to lowercase
+    fmts = tuple(export_formats()['Argument'][1:])  # --include arguments
+    flags = [x in include for x in fmts]
+    assert sum(flags) == len(include), f'ERROR: Invalid --include {include}, valid --include arguments are {fmts}'
+    jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle = flags  # export booleans
+    file = Path(url2file(weights) if str(weights).startswith(('http:/', 'https:/')) else weights)  # PyTorch weights
+
+    # Load PyTorch model
+    device = select_device(device)
+    if half:
+        assert device.type != 'cpu' or coreml, '--half only compatible with GPU export, i.e. use --device 0'
+        assert not dynamic, '--half not compatible with --dynamic, i.e. use either --half or --dynamic but not both'
+    model = attempt_load(weights, device=device, inplace=True, fuse=True)  # load FP32 model
+
+    # Checks
+    imgsz *= 2 if len(imgsz) == 1 else 1  # expand
+    if optimize:
+        assert device.type == 'cpu', '--optimize not compatible with cuda devices, i.e. use --device cpu'
+
+    # Input
+    gs = int(max(model.stride))  # grid size (max stride)
+    imgsz = [check_img_size(x, gs) for x in imgsz]  # verify img_size are gs-multiples
+    im = torch.zeros(batch_size, 3, *imgsz).to(device)  # image size(1,3,320,192) BCHW iDetection
+
+    # Update model
+    model.eval()
+    for k, m in model.named_modules():
+        if isinstance(m, (Detect, V6Detect)):
+            m.inplace = inplace
+            m.dynamic = dynamic
+            m.export = True
+
+    for _ in range(2):
+        y = model(im)  # dry runs
+    if half and not coreml:
+        im, model = im.half(), model.half()  # to FP16
+    shape = tuple((y[0] if isinstance(y, tuple) else y).shape)  # model output shape
+    metadata = {'stride': int(max(model.stride)), 'names': model.names}  # model metadata
+    LOGGER.info(f"\n{colorstr('PyTorch:')} starting from {file} with output shape {shape} ({file_size(file):.1f} MB)")
+
+    # Exports
+    f = [''] * len(fmts)  # exported filenames
+    warnings.filterwarnings(action='ignore', category=torch.jit.TracerWarning)  # suppress TracerWarning
+    if jit:  # TorchScript
+        f[0], _ = export_torchscript(model, im, file, optimize)
+    if engine:  # TensorRT required before ONNX
+        f[1], _ = export_engine(model, im, file, half, dynamic, simplify, workspace, verbose)
+    if onnx or xml:  # OpenVINO requires ONNX
+        f[2], _ = export_onnx(model, im, file, opset, dynamic, simplify)
+    if xml:  # OpenVINO
+        f[3], _ = export_openvino(file, metadata, half)
+    if coreml:  # CoreML
+        f[4], _ = export_coreml(model, im, file, int8, half)
+    if any((saved_model, pb, tflite, edgetpu, tfjs)):  # TensorFlow formats
+        assert not tflite or not tfjs, 'TFLite and TF.js models must be exported separately, please pass only one type.'
+        assert not isinstance(model, ClassificationModel), 'ClassificationModel export to TF formats not yet supported.'
+        f[5], s_model = export_saved_model(model.cpu(),
+                                           im,
+                                           file,
+                                           dynamic,
+                                           tf_nms=nms or agnostic_nms or tfjs,
+                                           agnostic_nms=agnostic_nms or tfjs,
+                                           topk_per_class=topk_per_class,
+                                           topk_all=topk_all,
+                                           iou_thres=iou_thres,
+                                           conf_thres=conf_thres,
+                                           keras=keras)
+        if pb or tfjs:  # pb prerequisite to tfjs
+            f[6], _ = export_pb(s_model, file)
+        if tflite or edgetpu:
+            f[7], _ = export_tflite(s_model, im, file, int8 or edgetpu, data=data, nms=nms, agnostic_nms=agnostic_nms)
+            if edgetpu:
+                f[8], _ = export_edgetpu(file)
+            add_tflite_metadata(f[8] or f[7], metadata, num_outputs=len(s_model.outputs))
+        if tfjs:
+            f[9], _ = export_tfjs(file)
+    if paddle:  # PaddlePaddle
+        f[10], _ = export_paddle(model, im, file, metadata)
+
+    # Finish
+    f = [str(x) for x in f if x]  # filter out '' and None
+    if any(f):
+        cls, det, seg = (isinstance(model, x) for x in (ClassificationModel, DetectionModel, SegmentationModel))  # type
+        dir = Path('segment' if seg else 'classify' if cls else '')
+        h = '--half' if half else ''  # --half FP16 inference arg
+        s = "# WARNING ⚠️ ClassificationModel not yet supported for PyTorch Hub AutoShape inference" if cls else \
+            "# WARNING ⚠️ SegmentationModel not yet supported for PyTorch Hub AutoShape inference" if seg else ''
+        LOGGER.info(f'\nExport complete ({time.time() - t:.1f}s)'
+                    f"\nResults saved to {colorstr('bold', file.parent.resolve())}"
+                    f"\nDetect:          python {dir / ('detect.py' if det else 'predict.py')} --weights {f[-1]} {h}"
+                    f"\nValidate:        python {dir / 'val.py'} --weights {f[-1]} {h}"
+                    f"\nPyTorch Hub:     model = torch.hub.load('ultralytics/yolov5', 'custom', '{f[-1]}')  {s}"
+                    f"\nVisualize:       https://netron.app")
+    return f  # return list of exported files/dirs
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo.pt', help='model.pt path(s)')
+    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640, 640], help='image (h, w)')
+    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
+    parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--half', action='store_true', help='FP16 half-precision export')
+    parser.add_argument('--inplace', action='store_true', help='set YOLO Detect() inplace=True')
+    parser.add_argument('--keras', action='store_true', help='TF: use Keras')
+    parser.add_argument('--optimize', action='store_true', help='TorchScript: optimize for mobile')
+    parser.add_argument('--int8', action='store_true', help='CoreML/TF INT8 quantization')
+    parser.add_argument('--dynamic', action='store_true', help='ONNX/TF/TensorRT: dynamic axes')
+    parser.add_argument('--simplify', action='store_true', help='ONNX: simplify model')
+    parser.add_argument('--opset', type=int, default=12, help='ONNX: opset version')
+    parser.add_argument('--verbose', action='store_true', help='TensorRT: verbose log')
+    parser.add_argument('--workspace', type=int, default=4, help='TensorRT: workspace size (GB)')
+    parser.add_argument('--nms', action='store_true', help='TF: add NMS to model')
+    parser.add_argument('--agnostic-nms', action='store_true', help='TF: add agnostic NMS to model')
+    parser.add_argument('--topk-per-class', type=int, default=100, help='TF.js NMS: topk per class to keep')
+    parser.add_argument('--topk-all', type=int, default=100, help='TF.js NMS: topk for all classes to keep')
+    parser.add_argument('--iou-thres', type=float, default=0.45, help='TF.js NMS: IoU threshold')
+    parser.add_argument('--conf-thres', type=float, default=0.25, help='TF.js NMS: confidence threshold')
+    parser.add_argument(
+        '--include',
+        nargs='+',
+        default=['torchscript'],
+        help='torchscript, onnx, openvino, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle')
+    opt = parser.parse_args()
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    for opt.weights in (opt.weights if isinstance(opt.weights, list) else [opt.weights]):
+        run(**vars(opt))
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

hubconf.py

@@ -0,0 +1,107 @@
+import torch
+
+
+def _create(name, pretrained=True, channels=3, classes=80, autoshape=True, verbose=True, device=None):
+    """Creates or loads a YOLO model
+
+    Arguments:
+        name (str): model name 'yolov3' or path 'path/to/best.pt'
+        pretrained (bool): load pretrained weights into the model
+        channels (int): number of input channels
+        classes (int): number of model classes
+        autoshape (bool): apply YOLO .autoshape() wrapper to model
+        verbose (bool): print all information to screen
+        device (str, torch.device, None): device to use for model parameters
+
+    Returns:
+        YOLO model
+    """
+    from pathlib import Path
+
+    from models.common import AutoShape, DetectMultiBackend
+    from models.experimental import attempt_load
+    from models.yolo import ClassificationModel, DetectionModel, SegmentationModel
+    from utils.downloads import attempt_download
+    from utils.general import LOGGER, check_requirements, intersect_dicts, logging
+    from utils.torch_utils import select_device
+
+    if not verbose:
+        LOGGER.setLevel(logging.WARNING)
+    check_requirements(exclude=('opencv-python', 'tensorboard', 'thop'))
+    name = Path(name)
+    path = name.with_suffix('.pt') if name.suffix == '' and not name.is_dir() else name  # checkpoint path
+    try:
+        device = select_device(device)
+        if pretrained and channels == 3 and classes == 80:
+            try:
+                model = DetectMultiBackend(path, device=device, fuse=autoshape)  # detection model
+                if autoshape:
+                    if model.pt and isinstance(model.model, ClassificationModel):
+                        LOGGER.warning('WARNING ⚠️ YOLO ClassificationModel is not yet AutoShape compatible. '
+                                       'You must pass torch tensors in BCHW to this model, i.e. shape(1,3,224,224).')
+                    elif model.pt and isinstance(model.model, SegmentationModel):
+                        LOGGER.warning('WARNING ⚠️ YOLO SegmentationModel is not yet AutoShape compatible. '
+                                       'You will not be able to run inference with this model.')
+                    else:
+                        model = AutoShape(model)  # for file/URI/PIL/cv2/np inputs and NMS
+            except Exception:
+                model = attempt_load(path, device=device, fuse=False)  # arbitrary model
+        else:
+            cfg = list((Path(__file__).parent / 'models').rglob(f'{path.stem}.yaml'))[0]  # model.yaml path
+            model = DetectionModel(cfg, channels, classes)  # create model
+            if pretrained:
+                ckpt = torch.load(attempt_download(path), map_location=device)  # load
+                csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+                csd = intersect_dicts(csd, model.state_dict(), exclude=['anchors'])  # intersect
+                model.load_state_dict(csd, strict=False)  # load
+                if len(ckpt['model'].names) == classes:
+                    model.names = ckpt['model'].names  # set class names attribute
+        if not verbose:
+            LOGGER.setLevel(logging.INFO)  # reset to default
+        return model.to(device)
+
+    except Exception as e:
+        help_url = 'https://github.com/ultralytics/yolov5/issues/36'
+        s = f'{e}. Cache may be out of date, try `force_reload=True` or see {help_url} for help.'
+        raise Exception(s) from e
+
+
+def custom(path='path/to/model.pt', autoshape=True, _verbose=True, device=None):
+    # YOLO custom or local model
+    return _create(path, autoshape=autoshape, verbose=_verbose, device=device)
+
+
+if __name__ == '__main__':
+    import argparse
+    from pathlib import Path
+
+    import numpy as np
+    from PIL import Image
+
+    from utils.general import cv2, print_args
+
+    # Argparser
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model', type=str, default='yolo', help='model name')
+    opt = parser.parse_args()
+    print_args(vars(opt))
+
+    # Model
+    model = _create(name=opt.model, pretrained=True, channels=3, classes=80, autoshape=True, verbose=True)
+    # model = custom(path='path/to/model.pt')  # custom
+
+    # Images
+    imgs = [
+        'data/images/zidane.jpg',  # filename
+        Path('data/images/zidane.jpg'),  # Path
+        'https://ultralytics.com/images/zidane.jpg',  # URI
+        cv2.imread('data/images/bus.jpg')[:, :, ::-1],  # OpenCV
+        Image.open('data/images/bus.jpg'),  # PIL
+        np.zeros((320, 640, 3))]  # numpy
+
+    # Inference
+    results = model(imgs, size=320)  # batched inference
+
+    # Results
+    results.print()
+    results.save()

train.py

@@ -0,0 +1,634 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import val as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import Model
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.dataloaders import create_dataloader
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_img_size,
+                           check_suffix, check_yaml, colorstr, get_latest_run, increment_path, init_seeds,
+                           intersect_dicts, labels_to_class_weights, labels_to_image_weights, methods,
+                           one_cycle, one_flat_cycle, print_args, print_mutation, strip_optimizer, yaml_save)
+from utils.loggers import Loggers
+from utils.loggers.comet.comet_utils import check_comet_resume
+from utils.loss_tal import ComputeLoss
+from utils.metrics import fitness
+from utils.plots import plot_evolve
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP,
+                               smart_optimizer, smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
+    callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+    last_striped, best_striped = w / 'last_striped.pt', w / 'best_striped.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    hyp['anchor_t'] = 5.0
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance
+
+        # Register actions
+        for k in methods(loggers):
+            callbacks.register_action(k, callback=getattr(loggers, k))
+
+        # Process custom dataset artifact link
+        data_dict = loggers.remote_dataset
+        if resume:  # If resuming runs from remote artifact
+            weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        # v.requires_grad = True  # train all layers TODO: uncomment this line as in master
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    elif opt.flat_cos_lr:
+        lf = one_flat_cycle(1, hyp['lrf'], epochs)  # flat cosine 1->hyp['lrf']        
+    elif opt.fixed_lr:
+        lf = lambda x: 1.0
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
+    # from utils.plots import plot_lr_scheduler; plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(train_path,
+                                              imgsz,
+                                              batch_size // WORLD_SIZE,
+                                              gs,
+                                              single_cls,
+                                              hyp=hyp,
+                                              augment=True,
+                                              cache=None if opt.cache == 'val' else opt.cache,
+                                              rect=opt.rect,
+                                              rank=LOCAL_RANK,
+                                              workers=workers,
+                                              image_weights=opt.image_weights,
+                                              close_mosaic=opt.close_mosaic != 0,
+                                              quad=opt.quad,
+                                              prefix=colorstr('train: '),
+                                              shuffle=True,
+                                              min_items=opt.min_items)
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            # if not opt.noautoanchor:
+            #     check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+        callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0)  # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model)  # init loss class
+    callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(3, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 7) % ('Epoch', 'GPU_mem', 'box_loss', 'cls_loss', 'dfl_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
+            callbacks.run('on_train_batch_start')
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 5) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths, list(mloss))
+                if callbacks.stop_training:
+                    return
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                results, maps, _ = validate.run(data_dict,
+                                                batch_size=batch_size // WORLD_SIZE * 2,
+                                                imgsz=imgsz,
+                                                half=amp,
+                                                model=ema.ema,
+                                                single_cls=single_cls,
+                                                dataloader=val_loader,
+                                                save_dir=save_dir,
+                                                plots=False,
+                                                callbacks=callbacks,
+                                                compute_loss=compute_loss)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, mAP@.5, mAP@.5-.95]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                del ckpt
+                callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                if f is last:
+                    strip_optimizer(f, last_striped)  # strip optimizers
+                else:
+                    strip_optimizer(f, best_striped)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss)  # val best model with plots
+                    if is_coco:
+                        callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+
+        callbacks.run('on_train_end', last, best, epoch, results)
+
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    # parser.add_argument('--weights', type=str, default=ROOT / 'yolo.pt', help='initial weights path')
+    # parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--weights', type=str, default='', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='yolo.yaml', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--flat-cos-lr', action='store_true', help='flat cosine LR scheduler')
+    parser.add_argument('--fixed-lr', action='store_true', help='fixed LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--min-items', type=int, default=0, help='Experimental')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Logger arguments
+    parser.add_argument('--entity', default=None, help='Entity')
+    parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='Upload data, "val" option')
+    parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval')
+    parser.add_argument('--artifact_alias', type=str, default='latest', help='Version of dataset artifact to use')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+
+    # Resume (from specified or most recent last.pt)
+    if opt.resume and not check_comet_resume(opt) and not opt.evolve:
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss',
+                    'val/obj_loss', 'val/cls_loss')
+            print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

train_dual.py

@@ -0,0 +1,644 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import val_dual as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import Model
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.dataloaders import create_dataloader
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_git_info,
+                           check_git_status, check_img_size, check_requirements, check_suffix, check_yaml, colorstr,
+                           get_latest_run, increment_path, init_seeds, intersect_dicts, labels_to_class_weights,
+                           labels_to_image_weights, methods, one_cycle, print_args, print_mutation, strip_optimizer,
+                           yaml_save, one_flat_cycle)
+from utils.loggers import Loggers
+from utils.loggers.comet.comet_utils import check_comet_resume
+from utils.loss_tal_dual import ComputeLoss
+#from utils.loss_tal_dual import ComputeLossLH as ComputeLoss
+#from utils.loss_tal_dual import ComputeLossLHCF as ComputeLoss
+from utils.metrics import fitness
+from utils.plots import plot_evolve
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP, smart_optimizer,
+                               smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None#check_git_info()
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
+    callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    hyp['anchor_t'] = 5.0
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance
+
+        # Register actions
+        for k in methods(loggers):
+            callbacks.register_action(k, callback=getattr(loggers, k))
+
+        # Process custom dataset artifact link
+        data_dict = loggers.remote_dataset
+        if resume:  # If resuming runs from remote artifact
+            weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        # v.requires_grad = True  # train all layers TODO: uncomment this line as in master
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    elif opt.flat_cos_lr:
+        lf = one_flat_cycle(1, hyp['lrf'], epochs)  # flat cosine 1->hyp['lrf']        
+    elif opt.fixed_lr:
+        lf = lambda x: 1.0
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+
+    # def lf(x):  # saw
+    #     return (1 - (x % 30) / 30) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+    #
+    # def lf(x):  # triangle start at min
+    #     return 2 * abs(x / 30 - math.floor(x / 30 + 1 / 2)) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+    #
+    # def lf(x):  # triangle start at max
+    #     return 2 * abs(x / 32 + .5 - math.floor(x / 32 + 1)) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
+    # from utils.plots import plot_lr_scheduler; plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(train_path,
+                                              imgsz,
+                                              batch_size // WORLD_SIZE,
+                                              gs,
+                                              single_cls,
+                                              hyp=hyp,
+                                              augment=True,
+                                              cache=None if opt.cache == 'val' else opt.cache,
+                                              rect=opt.rect,
+                                              rank=LOCAL_RANK,
+                                              workers=workers,
+                                              image_weights=opt.image_weights,
+                                              close_mosaic=opt.close_mosaic != 0,
+                                              quad=opt.quad,
+                                              prefix=colorstr('train: '),
+                                              shuffle=True,
+                                              min_items=opt.min_items)
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            # if not opt.noautoanchor:
+            #     check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+        callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0)  # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model)  # init loss class
+    callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(3, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 7) % ('Epoch', 'GPU_mem', 'box_loss', 'cls_loss', 'dfl_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
+            callbacks.run('on_train_batch_start')
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 5) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths, list(mloss))
+                if callbacks.stop_training:
+                    return
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                results, maps, _ = validate.run(data_dict,
+                                                batch_size=batch_size // WORLD_SIZE * 2,
+                                                imgsz=imgsz,
+                                                half=amp,
+                                                model=ema.ema,
+                                                single_cls=single_cls,
+                                                dataloader=val_loader,
+                                                save_dir=save_dir,
+                                                plots=False,
+                                                callbacks=callbacks,
+                                                compute_loss=compute_loss)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, mAP@.5, mAP@.5-.95]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                del ckpt
+                callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                strip_optimizer(f)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss)  # val best model with plots
+                    if is_coco:
+                        callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+
+        callbacks.run('on_train_end', last, best, epoch, results)
+
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    # parser.add_argument('--weights', type=str, default=ROOT / 'yolo.pt', help='initial weights path')
+    # parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--weights', type=str, default='', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='yolo.yaml', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-high.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--flat-cos-lr', action='store_true', help='flat cosine LR scheduler')
+    parser.add_argument('--fixed-lr', action='store_true', help='fixed LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--min-items', type=int, default=0, help='Experimental')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Logger arguments
+    parser.add_argument('--entity', default=None, help='Entity')
+    parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='Upload data, "val" option')
+    parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval')
+    parser.add_argument('--artifact_alias', type=str, default='latest', help='Version of dataset artifact to use')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+        #check_git_status()
+        #check_requirements()
+
+    # Resume (from specified or most recent last.pt)
+    if opt.resume and not check_comet_resume(opt) and not opt.evolve:
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss',
+                    'val/obj_loss', 'val/cls_loss')
+            print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

train_triple.py

@@ -0,0 +1,636 @@
+import argparse
+import math
+import os
+import random
+import sys
+import time
+from copy import deepcopy
+from datetime import datetime
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import yaml
+from torch.optim import lr_scheduler
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+import val_triple as validate  # for end-of-epoch mAP
+from models.experimental import attempt_load
+from models.yolo import Model
+from utils.autoanchor import check_anchors
+from utils.autobatch import check_train_batch_size
+from utils.callbacks import Callbacks
+from utils.dataloaders import create_dataloader
+from utils.downloads import attempt_download, is_url
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, check_amp, check_dataset, check_file, check_git_info,
+                           check_git_status, check_img_size, check_requirements, check_suffix, check_yaml, colorstr,
+                           get_latest_run, increment_path, init_seeds, intersect_dicts, labels_to_class_weights,
+                           labels_to_image_weights, methods, one_cycle, print_args, print_mutation, strip_optimizer,
+                           yaml_save)
+from utils.loggers import Loggers
+from utils.loggers.comet.comet_utils import check_comet_resume
+from utils.loss_tal_triple import ComputeLoss
+from utils.metrics import fitness
+from utils.plots import plot_evolve
+from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP, smart_optimizer,
+                               smart_resume, torch_distributed_zero_first)
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+GIT_INFO = None#check_git_info()
+
+
+def train(hyp, opt, device, callbacks):  # hyp is path/to/hyp.yaml or hyp dictionary
+    save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \
+        Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \
+        opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze
+    callbacks.run('on_pretrain_routine_start')
+
+    # Directories
+    w = save_dir / 'weights'  # weights dir
+    (w.parent if evolve else w).mkdir(parents=True, exist_ok=True)  # make dir
+    last, best = w / 'last.pt', w / 'best.pt'
+
+    # Hyperparameters
+    if isinstance(hyp, str):
+        with open(hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+    LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
+    hyp['anchor_t'] = 5.0
+    opt.hyp = hyp.copy()  # for saving hyps to checkpoints
+
+    # Save run settings
+    if not evolve:
+        yaml_save(save_dir / 'hyp.yaml', hyp)
+        yaml_save(save_dir / 'opt.yaml', vars(opt))
+
+    # Loggers
+    data_dict = None
+    if RANK in {-1, 0}:
+        loggers = Loggers(save_dir, weights, opt, hyp, LOGGER)  # loggers instance
+
+        # Register actions
+        for k in methods(loggers):
+            callbacks.register_action(k, callback=getattr(loggers, k))
+
+        # Process custom dataset artifact link
+        data_dict = loggers.remote_dataset
+        if resume:  # If resuming runs from remote artifact
+            weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size
+
+    # Config
+    plots = not evolve and not opt.noplots  # create plots
+    cuda = device.type != 'cpu'
+    init_seeds(opt.seed + 1 + RANK, deterministic=True)
+    with torch_distributed_zero_first(LOCAL_RANK):
+        data_dict = data_dict or check_dataset(data)  # check if None
+    train_path, val_path = data_dict['train'], data_dict['val']
+    nc = 1 if single_cls else int(data_dict['nc'])  # number of classes
+    names = {0: 'item'} if single_cls and len(data_dict['names']) != 1 else data_dict['names']  # class names
+    #is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt')  # COCO dataset
+    is_coco = isinstance(val_path, str) and val_path.endswith('val2017.txt')  # COCO dataset
+
+    # Model
+    check_suffix(weights, '.pt')  # check weights
+    pretrained = weights.endswith('.pt')
+    if pretrained:
+        with torch_distributed_zero_first(LOCAL_RANK):
+            weights = attempt_download(weights)  # download if not found locally
+        ckpt = torch.load(weights, map_location='cpu')  # load checkpoint to CPU to avoid CUDA memory leak
+        model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+        exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else []  # exclude keys
+        csd = ckpt['model'].float().state_dict()  # checkpoint state_dict as FP32
+        csd = intersect_dicts(csd, model.state_dict(), exclude=exclude)  # intersect
+        model.load_state_dict(csd, strict=False)  # load
+        LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}')  # report
+    else:
+        model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device)  # create
+    amp = check_amp(model)  # check AMP
+
+    # Freeze
+    freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))]  # layers to freeze
+    for k, v in model.named_parameters():
+        # v.requires_grad = True  # train all layers TODO: uncomment this line as in master
+        # v.register_hook(lambda x: torch.nan_to_num(x))  # NaN to 0 (commented for erratic training results)
+        if any(x in k for x in freeze):
+            LOGGER.info(f'freezing {k}')
+            v.requires_grad = False
+
+    # Image size
+    gs = max(int(model.stride.max()), 32)  # grid size (max stride)
+    imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2)  # verify imgsz is gs-multiple
+
+    # Batch size
+    if RANK == -1 and batch_size == -1:  # single-GPU only, estimate best batch size
+        batch_size = check_train_batch_size(model, imgsz, amp)
+        loggers.on_params_update({"batch_size": batch_size})
+
+    # Optimizer
+    nbs = 64  # nominal batch size
+    accumulate = max(round(nbs / batch_size), 1)  # accumulate loss before optimizing
+    hyp['weight_decay'] *= batch_size * accumulate / nbs  # scale weight_decay
+    optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay'])
+
+    # Scheduler
+    if opt.cos_lr:
+        lf = one_cycle(1, hyp['lrf'], epochs)  # cosine 1->hyp['lrf']
+    else:
+        lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']  # linear
+
+    # def lf(x):  # saw
+    #     return (1 - (x % 30) / 30) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+    #
+    # def lf(x):  # triangle start at min
+    #     return 2 * abs(x / 30 - math.floor(x / 30 + 1 / 2)) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+    #
+    # def lf(x):  # triangle start at max
+    #     return 2 * abs(x / 32 + .5 - math.floor(x / 32 + 1)) * (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf']
+
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
+    # from utils.plots import plot_lr_scheduler; plot_lr_scheduler(optimizer, scheduler, epochs)
+
+    # EMA
+    ema = ModelEMA(model) if RANK in {-1, 0} else None
+
+    # Resume
+    best_fitness, start_epoch = 0.0, 0
+    if pretrained:
+        if resume:
+            best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume)
+        del ckpt, csd
+
+    # DP mode
+    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
+        LOGGER.warning('WARNING ⚠️ DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.')
+        model = torch.nn.DataParallel(model)
+
+    # SyncBatchNorm
+    if opt.sync_bn and cuda and RANK != -1:
+        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device)
+        LOGGER.info('Using SyncBatchNorm()')
+
+    # Trainloader
+    train_loader, dataset = create_dataloader(train_path,
+                                              imgsz,
+                                              batch_size // WORLD_SIZE,
+                                              gs,
+                                              single_cls,
+                                              hyp=hyp,
+                                              augment=True,
+                                              cache=None if opt.cache == 'val' else opt.cache,
+                                              rect=opt.rect,
+                                              rank=LOCAL_RANK,
+                                              workers=workers,
+                                              image_weights=opt.image_weights,
+                                              close_mosaic=opt.close_mosaic != 0,
+                                              quad=opt.quad,
+                                              prefix=colorstr('train: '),
+                                              shuffle=True,
+                                              min_items=opt.min_items)
+    labels = np.concatenate(dataset.labels, 0)
+    mlc = int(labels[:, 0].max())  # max label class
+    assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}'
+
+    # Process 0
+    if RANK in {-1, 0}:
+        val_loader = create_dataloader(val_path,
+                                       imgsz,
+                                       batch_size // WORLD_SIZE * 2,
+                                       gs,
+                                       single_cls,
+                                       hyp=hyp,
+                                       cache=None if noval else opt.cache,
+                                       rect=True,
+                                       rank=-1,
+                                       workers=workers * 2,
+                                       pad=0.5,
+                                       prefix=colorstr('val: '))[0]
+
+        if not resume:
+            # if not opt.noautoanchor:
+            #     check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz)  # run AutoAnchor
+            model.half().float()  # pre-reduce anchor precision
+
+        callbacks.run('on_pretrain_routine_end', labels, names)
+
+    # DDP mode
+    if cuda and RANK != -1:
+        model = smart_DDP(model)
+
+    # Model attributes
+    nl = de_parallel(model).model[-1].nl  # number of detection layers (to scale hyps)
+    #hyp['box'] *= 3 / nl  # scale to layers
+    #hyp['cls'] *= nc / 80 * 3 / nl  # scale to classes and layers
+    #hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl  # scale to image size and layers
+    hyp['label_smoothing'] = opt.label_smoothing
+    model.nc = nc  # attach number of classes to model
+    model.hyp = hyp  # attach hyperparameters to model
+    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights
+    model.names = names
+
+    # Start training
+    t0 = time.time()
+    nb = len(train_loader)  # number of batches
+    nw = max(round(hyp['warmup_epochs'] * nb), 100)  # number of warmup iterations, max(3 epochs, 100 iterations)
+    # nw = min(nw, (epochs - start_epoch) / 2 * nb)  # limit warmup to < 1/2 of training
+    last_opt_step = -1
+    maps = np.zeros(nc)  # mAP per class
+    results = (0, 0, 0, 0, 0, 0, 0)  # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls)
+    scheduler.last_epoch = start_epoch - 1  # do not move
+    scaler = torch.cuda.amp.GradScaler(enabled=amp)
+    stopper, stop = EarlyStopping(patience=opt.patience), False
+    compute_loss = ComputeLoss(model)  # init loss class
+    callbacks.run('on_train_start')
+    LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n'
+                f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n'
+                f"Logging results to {colorstr('bold', save_dir)}\n"
+                f'Starting training for {epochs} epochs...')
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        callbacks.run('on_train_epoch_start')
+        model.train()
+
+        # Update image weights (optional, single-GPU only)
+        if opt.image_weights:
+            cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc  # class weights
+            iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw)  # image weights
+            dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n)  # rand weighted idx
+        if epoch == (epochs - opt.close_mosaic):
+            LOGGER.info("Closing dataloader mosaic")
+            dataset.mosaic = False
+
+        # Update mosaic border (optional)
+        # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs)
+        # dataset.mosaic_border = [b - imgsz, -b]  # height, width borders
+
+        mloss = torch.zeros(3, device=device)  # mean losses
+        if RANK != -1:
+            train_loader.sampler.set_epoch(epoch)
+        pbar = enumerate(train_loader)
+        LOGGER.info(('\n' + '%11s' * 7) % ('Epoch', 'GPU_mem', 'box_loss', 'cls_loss', 'dfl_loss', 'Instances', 'Size'))
+        if RANK in {-1, 0}:
+            pbar = tqdm(pbar, total=nb, bar_format=TQDM_BAR_FORMAT)  # progress bar
+        optimizer.zero_grad()
+        for i, (imgs, targets, paths, _) in pbar:  # batch -------------------------------------------------------------
+            callbacks.run('on_train_batch_start')
+            ni = i + nb * epoch  # number integrated batches (since train start)
+            imgs = imgs.to(device, non_blocking=True).float() / 255  # uint8 to float32, 0-255 to 0.0-1.0
+
+            # Warmup
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0])  # iou loss ratio (obj_loss = 1.0 or iou)
+                accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round())
+                for j, x in enumerate(optimizer.param_groups):
+                    # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
+                    x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']])
+
+            # Multi-scale
+            if opt.multi_scale:
+                sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs  # size
+                sf = sz / max(imgs.shape[2:])  # scale factor
+                if sf != 1:
+                    ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]]  # new shape (stretched to gs-multiple)
+                    imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
+
+            # Forward
+            with torch.cuda.amp.autocast(amp):
+                pred = model(imgs)  # forward
+                loss, loss_items = compute_loss(pred, targets.to(device))  # loss scaled by batch_size
+                if RANK != -1:
+                    loss *= WORLD_SIZE  # gradient averaged between devices in DDP mode
+                if opt.quad:
+                    loss *= 4.
+
+            # Backward
+            scaler.scale(loss).backward()
+
+            # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html
+            if ni - last_opt_step >= accumulate:
+                scaler.unscale_(optimizer)  # unscale gradients
+                torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0)  # clip gradients
+                scaler.step(optimizer)  # optimizer.step
+                scaler.update()
+                optimizer.zero_grad()
+                if ema:
+                    ema.update(model)
+                last_opt_step = ni
+
+            # Log
+            if RANK in {-1, 0}:
+                mloss = (mloss * i + loss_items) / (i + 1)  # update mean losses
+                mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G'  # (GB)
+                pbar.set_description(('%11s' * 2 + '%11.4g' * 5) %
+                                     (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1]))
+                callbacks.run('on_train_batch_end', model, ni, imgs, targets, paths, list(mloss))
+                if callbacks.stop_training:
+                    return
+            # end batch ------------------------------------------------------------------------------------------------
+
+        # Scheduler
+        lr = [x['lr'] for x in optimizer.param_groups]  # for loggers
+        scheduler.step()
+
+        if RANK in {-1, 0}:
+            # mAP
+            callbacks.run('on_train_epoch_end', epoch=epoch)
+            ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights'])
+            final_epoch = (epoch + 1 == epochs) or stopper.possible_stop
+            if not noval or final_epoch:  # Calculate mAP
+                results, maps, _ = validate.run(data_dict,
+                                                batch_size=batch_size // WORLD_SIZE * 2,
+                                                imgsz=imgsz,
+                                                half=amp,
+                                                model=ema.ema,
+                                                single_cls=single_cls,
+                                                dataloader=val_loader,
+                                                save_dir=save_dir,
+                                                plots=False,
+                                                callbacks=callbacks,
+                                                compute_loss=compute_loss)
+
+            # Update best mAP
+            fi = fitness(np.array(results).reshape(1, -1))  # weighted combination of [P, R, mAP@.5, mAP@.5-.95]
+            stop = stopper(epoch=epoch, fitness=fi)  # early stop check
+            if fi > best_fitness:
+                best_fitness = fi
+            log_vals = list(mloss) + list(results) + lr
+            callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi)
+
+            # Save model
+            if (not nosave) or (final_epoch and not evolve):  # if save
+                ckpt = {
+                    'epoch': epoch,
+                    'best_fitness': best_fitness,
+                    'model': deepcopy(de_parallel(model)).half(),
+                    'ema': deepcopy(ema.ema).half(),
+                    'updates': ema.updates,
+                    'optimizer': optimizer.state_dict(),
+                    'opt': vars(opt),
+                    'git': GIT_INFO,  # {remote, branch, commit} if a git repo
+                    'date': datetime.now().isoformat()}
+
+                # Save last, best and delete
+                torch.save(ckpt, last)
+                if best_fitness == fi:
+                    torch.save(ckpt, best)
+                if opt.save_period > 0 and epoch % opt.save_period == 0:
+                    torch.save(ckpt, w / f'epoch{epoch}.pt')
+                del ckpt
+                callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi)
+
+        # EarlyStopping
+        if RANK != -1:  # if DDP training
+            broadcast_list = [stop if RANK == 0 else None]
+            dist.broadcast_object_list(broadcast_list, 0)  # broadcast 'stop' to all ranks
+            if RANK != 0:
+                stop = broadcast_list[0]
+        if stop:
+            break  # must break all DDP ranks
+
+        # end epoch ----------------------------------------------------------------------------------------------------
+    # end training -----------------------------------------------------------------------------------------------------
+    if RANK in {-1, 0}:
+        LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.')
+        for f in last, best:
+            if f.exists():
+                strip_optimizer(f)  # strip optimizers
+                if f is best:
+                    LOGGER.info(f'\nValidating {f}...')
+                    results, _, _ = validate.run(
+                        data_dict,
+                        batch_size=batch_size // WORLD_SIZE * 2,
+                        imgsz=imgsz,
+                        model=attempt_load(f, device).half(),
+                        single_cls=single_cls,
+                        dataloader=val_loader,
+                        save_dir=save_dir,
+                        save_json=is_coco,
+                        verbose=True,
+                        plots=plots,
+                        callbacks=callbacks,
+                        compute_loss=compute_loss)  # val best model with plots
+                    if is_coco:
+                        callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi)
+
+        callbacks.run('on_train_end', last, best, epoch, results)
+
+    torch.cuda.empty_cache()
+    return results
+
+
+def parse_opt(known=False):
+    parser = argparse.ArgumentParser()
+    # parser.add_argument('--weights', type=str, default=ROOT / 'yolo.pt', help='initial weights path')
+    # parser.add_argument('--cfg', type=str, default='', help='model.yaml path')
+    parser.add_argument('--weights', type=str, default='', help='initial weights path')
+    parser.add_argument('--cfg', type=str, default='yolo.yaml', help='model.yaml path')
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco.yaml', help='dataset.yaml path')
+    parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-high.yaml', help='hyperparameters path')
+    parser.add_argument('--epochs', type=int, default=100, help='total training epochs')
+    parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)')
+    parser.add_argument('--rect', action='store_true', help='rectangular training')
+    parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training')
+    parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
+    parser.add_argument('--noval', action='store_true', help='only validate final epoch')
+    parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor')
+    parser.add_argument('--noplots', action='store_true', help='save no plot files')
+    parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations')
+    parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
+    parser.add_argument('--cache', type=str, nargs='?', const='ram', help='image --cache ram/disk')
+    parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%')
+    parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class')
+    parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW', 'LION'], default='SGD', help='optimizer')
+    parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--quad', action='store_true', help='quad dataloader')
+    parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler')
+    parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon')
+    parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)')
+    parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2')
+    parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)')
+    parser.add_argument('--seed', type=int, default=0, help='Global training seed')
+    parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify')
+    parser.add_argument('--min-items', type=int, default=0, help='Experimental')
+    parser.add_argument('--close-mosaic', type=int, default=0, help='Experimental')
+
+    # Logger arguments
+    parser.add_argument('--entity', default=None, help='Entity')
+    parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='Upload data, "val" option')
+    parser.add_argument('--bbox_interval', type=int, default=-1, help='Set bounding-box image logging interval')
+    parser.add_argument('--artifact_alias', type=str, default='latest', help='Version of dataset artifact to use')
+
+    return parser.parse_known_args()[0] if known else parser.parse_args()
+
+
+def main(opt, callbacks=Callbacks()):
+    # Checks
+    if RANK in {-1, 0}:
+        print_args(vars(opt))
+        #check_git_status()
+        #check_requirements()
+
+    # Resume (from specified or most recent last.pt)
+    if opt.resume and not check_comet_resume(opt) and not opt.evolve:
+        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
+        opt_yaml = last.parent.parent / 'opt.yaml'  # train options yaml
+        opt_data = opt.data  # original dataset
+        if opt_yaml.is_file():
+            with open(opt_yaml, errors='ignore') as f:
+                d = yaml.safe_load(f)
+        else:
+            d = torch.load(last, map_location='cpu')['opt']
+        opt = argparse.Namespace(**d)  # replace
+        opt.cfg, opt.weights, opt.resume = '', str(last), True  # reinstate
+        if is_url(opt_data):
+            opt.data = check_file(opt_data)  # avoid HUB resume auth timeout
+    else:
+        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \
+            check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project)  # checks
+        assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified'
+        if opt.evolve:
+            if opt.project == str(ROOT / 'runs/train'):  # if default project name, rename to runs/evolve
+                opt.project = str(ROOT / 'runs/evolve')
+            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
+        if opt.name == 'cfg':
+            opt.name = Path(opt.cfg).stem  # use model.yaml as name
+        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))
+
+    # DDP mode
+    device = select_device(opt.device, batch_size=opt.batch_size)
+    if LOCAL_RANK != -1:
+        msg = 'is not compatible with YOLO Multi-GPU DDP training'
+        assert not opt.image_weights, f'--image-weights {msg}'
+        assert not opt.evolve, f'--evolve {msg}'
+        assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size'
+        assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE'
+        assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command'
+        torch.cuda.set_device(LOCAL_RANK)
+        device = torch.device('cuda', LOCAL_RANK)
+        dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo")
+
+    # Train
+    if not opt.evolve:
+        train(opt.hyp, opt, device, callbacks)
+
+    # Evolve hyperparameters (optional)
+    else:
+        # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)
+        meta = {
+            'lr0': (1, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
+            'lrf': (1, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
+            'momentum': (0.3, 0.6, 0.98),  # SGD momentum/Adam beta1
+            'weight_decay': (1, 0.0, 0.001),  # optimizer weight decay
+            'warmup_epochs': (1, 0.0, 5.0),  # warmup epochs (fractions ok)
+            'warmup_momentum': (1, 0.0, 0.95),  # warmup initial momentum
+            'warmup_bias_lr': (1, 0.0, 0.2),  # warmup initial bias lr
+            'box': (1, 0.02, 0.2),  # box loss gain
+            'cls': (1, 0.2, 4.0),  # cls loss gain
+            'cls_pw': (1, 0.5, 2.0),  # cls BCELoss positive_weight
+            'obj': (1, 0.2, 4.0),  # obj loss gain (scale with pixels)
+            'obj_pw': (1, 0.5, 2.0),  # obj BCELoss positive_weight
+            'iou_t': (0, 0.1, 0.7),  # IoU training threshold
+            'anchor_t': (1, 2.0, 8.0),  # anchor-multiple threshold
+            'anchors': (2, 2.0, 10.0),  # anchors per output grid (0 to ignore)
+            'fl_gamma': (0, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
+            'hsv_h': (1, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
+            'hsv_s': (1, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
+            'hsv_v': (1, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
+            'degrees': (1, 0.0, 45.0),  # image rotation (+/- deg)
+            'translate': (1, 0.0, 0.9),  # image translation (+/- fraction)
+            'scale': (1, 0.0, 0.9),  # image scale (+/- gain)
+            'shear': (1, 0.0, 10.0),  # image shear (+/- deg)
+            'perspective': (0, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
+            'flipud': (1, 0.0, 1.0),  # image flip up-down (probability)
+            'fliplr': (0, 0.0, 1.0),  # image flip left-right (probability)
+            'mosaic': (1, 0.0, 1.0),  # image mixup (probability)
+            'mixup': (1, 0.0, 1.0),  # image mixup (probability)
+            'copy_paste': (1, 0.0, 1.0)}  # segment copy-paste (probability)
+
+        with open(opt.hyp, errors='ignore') as f:
+            hyp = yaml.safe_load(f)  # load hyps dict
+            if 'anchors' not in hyp:  # anchors commented in hyp.yaml
+                hyp['anchors'] = 3
+        if opt.noautoanchor:
+            del hyp['anchors'], meta['anchors']
+        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
+        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
+        evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv'
+        if opt.bucket:
+            os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}')  # download evolve.csv if exists
+
+        for _ in range(opt.evolve):  # generations to evolve
+            if evolve_csv.exists():  # if evolve.csv exists: select best hyps and mutate
+                # Select parent(s)
+                parent = 'single'  # parent selection method: 'single' or 'weighted'
+                x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1)
+                n = min(5, len(x))  # number of previous results to consider
+                x = x[np.argsort(-fitness(x))][:n]  # top n mutations
+                w = fitness(x) - fitness(x).min() + 1E-6  # weights (sum > 0)
+                if parent == 'single' or len(x) == 1:
+                    # x = x[random.randint(0, n - 1)]  # random selection
+                    x = x[random.choices(range(n), weights=w)[0]]  # weighted selection
+                elif parent == 'weighted':
+                    x = (x * w.reshape(n, 1)).sum(0) / w.sum()  # weighted combination
+
+                # Mutate
+                mp, s = 0.8, 0.2  # mutation probability, sigma
+                npr = np.random
+                npr.seed(int(time.time()))
+                g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
+                ng = len(meta)
+                v = np.ones(ng)
+                while all(v == 1):  # mutate until a change occurs (prevent duplicates)
+                    v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0)
+                for i, k in enumerate(hyp.keys()):  # plt.hist(v.ravel(), 300)
+                    hyp[k] = float(x[i + 7] * v[i])  # mutate
+
+            # Constrain to limits
+            for k, v in meta.items():
+                hyp[k] = max(hyp[k], v[1])  # lower limit
+                hyp[k] = min(hyp[k], v[2])  # upper limit
+                hyp[k] = round(hyp[k], 5)  # significant digits
+
+            # Train mutation
+            results = train(hyp.copy(), opt, device, callbacks)
+            callbacks = Callbacks()
+            # Write mutation results
+            keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss',
+                    'val/obj_loss', 'val/cls_loss')
+            print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)
+
+        # Plot results
+        plot_evolve(evolve_csv)
+        LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n'
+                    f"Results saved to {colorstr('bold', save_dir)}\n"
+                    f'Usage example: $ python train.py --hyp {evolve_yaml}')
+
+
+def run(**kwargs):
+    # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolo.pt')
+    opt = parse_opt(True)
+    for k, v in kwargs.items():
+        setattr(opt, k, v)
+    main(opt)
+    return opt
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

val.py

@@ -0,0 +1,389 @@
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.callbacks import Callbacks
+from utils.dataloaders import create_dataloader
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, Profile, check_dataset, check_img_size, check_requirements,
+                           check_yaml, coco80_to_coco91_class, colorstr, increment_path, non_max_suppression,
+                           print_args, scale_boxes, xywh2xyxy, xyxy2xywh)
+from utils.metrics import ConfusionMatrix, ap_per_class, box_iou
+from utils.plots import output_to_target, plot_images, plot_val_study
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+def save_one_txt(predn, save_conf, shape, file):
+    # Save one txt result
+    gn = torch.tensor(shape)[[1, 0, 1, 0]]  # normalization gain whwh
+    for *xyxy, conf, cls in predn.tolist():
+        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
+        line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
+        with open(file, 'a') as f:
+            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+
+def save_one_json(predn, jdict, path, class_map):
+    # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}
+    image_id = int(path.stem) if path.stem.isnumeric() else path.stem
+    box = xyxy2xywh(predn[:, :4])  # xywh
+    box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
+    for p, b in zip(predn.tolist(), box.tolist()):
+        jdict.append({
+            'image_id': image_id,
+            'category_id': class_map[int(p[5])],
+            'bbox': [round(x, 3) for x in b],
+            'score': round(p[4], 5)})
+
+
+def process_batch(detections, labels, iouv):
+    """
+    Return correct prediction matrix
+    Arguments:
+        detections (array[N, 6]), x1, y1, x2, y2, conf, class
+        labels (array[M, 5]), class, x1, y1, x2, y2
+    Returns:
+        correct (array[N, 10]), for 10 IoU levels
+    """
+    correct = np.zeros((detections.shape[0], iouv.shape[0])).astype(bool)
+    iou = box_iou(labels[:, 1:], detections[:, :4])
+    correct_class = labels[:, 0:1] == detections[:, 5]
+    for i in range(len(iouv)):
+        x = torch.where((iou >= iouv[i]) & correct_class)  # IoU > threshold and classes match
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()  # [label, detect, iou]
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                # matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+            correct[matches[:, 1].astype(int), i] = True
+    return torch.tensor(correct, dtype=torch.bool, device=iouv.device)
+
+
+@smart_inference_mode()
+def run(
+        data,
+        weights=None,  # model.pt path(s)
+        batch_size=32,  # batch size
+        imgsz=640,  # inference size (pixels)
+        conf_thres=0.001,  # confidence threshold
+        iou_thres=0.7,  # NMS IoU threshold
+        max_det=300,  # maximum detections per image
+        task='val',  # train, val, test, speed or study
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        workers=8,  # max dataloader workers (per RANK in DDP mode)
+        single_cls=False,  # treat as single-class dataset
+        augment=False,  # augmented inference
+        verbose=False,  # verbose output
+        save_txt=False,  # save results to *.txt
+        save_hybrid=False,  # save label+prediction hybrid results to *.txt
+        save_conf=False,  # save confidences in --save-txt labels
+        save_json=False,  # save a COCO-JSON results file
+        project=ROOT / 'runs/val',  # save to project/name
+        name='exp',  # save to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        half=True,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        min_items=0,  # Experimental
+        model=None,
+        dataloader=None,
+        save_dir=Path(''),
+        plots=True,
+        callbacks=Callbacks(),
+        compute_loss=None,
+):
+    # Initialize/load model and set device
+    training = model is not None
+    if training:  # called by train.py
+        device, pt, jit, engine = next(model.parameters()).device, True, False, False  # get model device, PyTorch model
+        half &= device.type != 'cpu'  # half precision only supported on CUDA
+        model.half() if half else model.float()
+    else:  # called directly
+        device = select_device(device, batch_size=batch_size)
+
+        # Directories
+        save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+        (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+        # Load model
+        model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+        stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine
+        imgsz = check_img_size(imgsz, s=stride)  # check image size
+        half = model.fp16  # FP16 supported on limited backends with CUDA
+        if engine:
+            batch_size = model.batch_size
+        else:
+            device = model.device
+            if not (pt or jit):
+                batch_size = 1  # export.py models default to batch-size 1
+                LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models')
+
+        # Data
+        data = check_dataset(data)  # check
+
+    # Configure
+    model.eval()
+    cuda = device.type != 'cpu'
+    #is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'coco{os.sep}val2017.txt')  # COCO dataset
+    is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'val2017.txt')  # COCO dataset
+    nc = 1 if single_cls else int(data['nc'])  # number of classes
+    iouv = torch.linspace(0.5, 0.95, 10, device=device)  # iou vector for mAP@0.5:0.95
+    niou = iouv.numel()
+
+    # Dataloader
+    if not training:
+        if pt and not single_cls:  # check --weights are trained on --data
+            ncm = model.model.nc
+            assert ncm == nc, f'{weights} ({ncm} classes) trained on different --data than what you passed ({nc} ' \
+                              f'classes). Pass correct combination of --weights and --data that are trained together.'
+        model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz))  # warmup
+        pad, rect = (0.0, False) if task == 'speed' else (0.5, pt)  # square inference for benchmarks
+        task = task if task in ('train', 'val', 'test') else 'val'  # path to train/val/test images
+        dataloader = create_dataloader(data[task],
+                                       imgsz,
+                                       batch_size,
+                                       stride,
+                                       single_cls,
+                                       pad=pad,
+                                       rect=rect,
+                                       workers=workers,
+                                       min_items=opt.min_items,
+                                       prefix=colorstr(f'{task}: '))[0]
+
+    seen = 0
+    confusion_matrix = ConfusionMatrix(nc=nc)
+    names = model.names if hasattr(model, 'names') else model.module.names  # get class names
+    if isinstance(names, (list, tuple)):  # old format
+        names = dict(enumerate(names))
+    class_map = coco80_to_coco91_class() if is_coco else list(range(1000))
+    s = ('%22s' + '%11s' * 6) % ('Class', 'Images', 'Instances', 'P', 'R', 'mAP50', 'mAP50-95')
+    tp, fp, p, r, f1, mp, mr, map50, ap50, map = 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
+    dt = Profile(), Profile(), Profile()  # profiling times
+    loss = torch.zeros(3, device=device)
+    jdict, stats, ap, ap_class = [], [], [], []
+    callbacks.run('on_val_start')
+    pbar = tqdm(dataloader, desc=s, bar_format=TQDM_BAR_FORMAT)  # progress bar
+    for batch_i, (im, targets, paths, shapes) in enumerate(pbar):
+        callbacks.run('on_val_batch_start')
+        with dt[0]:
+            if cuda:
+                im = im.to(device, non_blocking=True)
+                targets = targets.to(device)
+            im = im.half() if half else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            nb, _, height, width = im.shape  # batch size, channels, height, width
+
+        # Inference
+        with dt[1]:
+            preds, train_out = model(im) if compute_loss else (model(im, augment=augment), None)
+
+        # Loss
+        if compute_loss:
+            loss += compute_loss(train_out, targets)[1]  # box, obj, cls
+
+        # NMS
+        targets[:, 2:] *= torch.tensor((width, height, width, height), device=device)  # to pixels
+        lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else []  # for autolabelling
+        with dt[2]:
+            preds = non_max_suppression(preds,
+                                        conf_thres,
+                                        iou_thres,
+                                        labels=lb,
+                                        multi_label=True,
+                                        agnostic=single_cls,
+                                        max_det=max_det)
+
+        # Metrics
+        for si, pred in enumerate(preds):
+            labels = targets[targets[:, 0] == si, 1:]
+            nl, npr = labels.shape[0], pred.shape[0]  # number of labels, predictions
+            path, shape = Path(paths[si]), shapes[si][0]
+            correct = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            seen += 1
+
+            if npr == 0:
+                if nl:
+                    stats.append((correct, *torch.zeros((2, 0), device=device), labels[:, 0]))
+                    if plots:
+                        confusion_matrix.process_batch(detections=None, labels=labels[:, 0])
+                continue
+
+            # Predictions
+            if single_cls:
+                pred[:, 5] = 0
+            predn = pred.clone()
+            scale_boxes(im[si].shape[1:], predn[:, :4], shape, shapes[si][1])  # native-space pred
+
+            # Evaluate
+            if nl:
+                tbox = xywh2xyxy(labels[:, 1:5])  # target boxes
+                scale_boxes(im[si].shape[1:], tbox, shape, shapes[si][1])  # native-space labels
+                labelsn = torch.cat((labels[:, 0:1], tbox), 1)  # native-space labels
+                correct = process_batch(predn, labelsn, iouv)
+                if plots:
+                    confusion_matrix.process_batch(predn, labelsn)
+            stats.append((correct, pred[:, 4], pred[:, 5], labels[:, 0]))  # (correct, conf, pcls, tcls)
+
+            # Save/log
+            if save_txt:
+                save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / f'{path.stem}.txt')
+            if save_json:
+                save_one_json(predn, jdict, path, class_map)  # append to COCO-JSON dictionary
+            callbacks.run('on_val_image_end', pred, predn, path, names, im[si])
+
+        # Plot images
+        if plots and batch_i < 3:
+            plot_images(im, targets, paths, save_dir / f'val_batch{batch_i}_labels.jpg', names)  # labels
+            plot_images(im, output_to_target(preds), paths, save_dir / f'val_batch{batch_i}_pred.jpg', names)  # pred
+
+        callbacks.run('on_val_batch_end', batch_i, im, targets, paths, shapes, preds)
+
+    # Compute metrics
+    stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)]  # to numpy
+    if len(stats) and stats[0].any():
+        tp, fp, p, r, f1, ap, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names)
+        ap50, ap = ap[:, 0], ap.mean(1)  # AP@0.5, AP@0.5:0.95
+        mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean()
+    nt = np.bincount(stats[3].astype(int), minlength=nc)  # number of targets per class
+
+    # Print results
+    pf = '%22s' + '%11i' * 2 + '%11.3g' * 4  # print format
+    LOGGER.info(pf % ('all', seen, nt.sum(), mp, mr, map50, map))
+    if nt.sum() == 0:
+        LOGGER.warning(f'WARNING ⚠️ no labels found in {task} set, can not compute metrics without labels')
+
+    # Print results per class
+    if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats):
+        for i, c in enumerate(ap_class):
+            LOGGER.info(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i]))
+
+    # Print speeds
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    if not training:
+        shape = (batch_size, 3, imgsz, imgsz)
+        LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t)
+
+    # Plots
+    if plots:
+        confusion_matrix.plot(save_dir=save_dir, names=list(names.values()))
+        callbacks.run('on_val_end', nt, tp, fp, p, r, f1, ap, ap50, ap_class, confusion_matrix)
+
+    # Save JSON
+    if save_json and len(jdict):
+        w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else ''  # weights
+        anno_json = str(Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json')  # annotations json
+        pred_json = str(save_dir / f"{w}_predictions.json")  # predictions json
+        LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...')
+        with open(pred_json, 'w') as f:
+            json.dump(jdict, f)
+
+        try:  # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
+            check_requirements('pycocotools')
+            from pycocotools.coco import COCO
+            from pycocotools.cocoeval import COCOeval
+
+            anno = COCO(anno_json)  # init annotations api
+            pred = anno.loadRes(pred_json)  # init predictions api
+            eval = COCOeval(anno, pred, 'bbox')
+            if is_coco:
+                eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files]  # image IDs to evaluate
+            eval.evaluate()
+            eval.accumulate()
+            eval.summarize()
+            map, map50 = eval.stats[:2]  # update results (mAP@0.5:0.95, mAP@0.5)
+        except Exception as e:
+            LOGGER.info(f'pycocotools unable to run: {e}')
+
+    # Return results
+    model.float()  # for training
+    if not training:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    maps = np.zeros(nc) + map
+    for i, c in enumerate(ap_class):
+        maps[c] = ap[i]
+    return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo.pt', help='model path(s)')
+    parser.add_argument('--batch-size', type=int, default=32, help='batch size')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.7, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=300, help='maximum detections per image')
+    parser.add_argument('--task', default='val', help='train, val, test, speed or study')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--verbose', action='store_true', help='report mAP by class')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-json', action='store_true', help='save a COCO-JSON results file')
+    parser.add_argument('--project', default=ROOT / 'runs/val', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    parser.add_argument('--min-items', type=int, default=0, help='Experimental')
+    opt = parser.parse_args()
+    opt.data = check_yaml(opt.data)  # check YAML
+    opt.save_json |= opt.data.endswith('coco.yaml')
+    opt.save_txt |= opt.save_hybrid
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    #check_requirements(exclude=('tensorboard', 'thop'))
+
+    if opt.task in ('train', 'val', 'test'):  # run normally
+        if opt.conf_thres > 0.001:  # https://github.com/ultralytics/yolov5/issues/1466
+            LOGGER.info(f'WARNING ⚠️ confidence threshold {opt.conf_thres} > 0.001 produces invalid results')
+        if opt.save_hybrid:
+            LOGGER.info('WARNING ⚠️ --save-hybrid will return high mAP from hybrid labels, not from predictions alone')
+        run(**vars(opt))
+
+    else:
+        weights = opt.weights if isinstance(opt.weights, list) else [opt.weights]
+        opt.half = torch.cuda.is_available() and opt.device != 'cpu'  # FP16 for fastest results
+        if opt.task == 'speed':  # speed benchmarks
+            # python val.py --task speed --data coco.yaml --batch 1 --weights yolo.pt...
+            opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False
+            for opt.weights in weights:
+                run(**vars(opt), plots=False)
+
+        elif opt.task == 'study':  # speed vs mAP benchmarks
+            # python val.py --task study --data coco.yaml --iou 0.7 --weights yolo.pt...
+            for opt.weights in weights:
+                f = f'study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt'  # filename to save to
+                x, y = list(range(256, 1536 + 128, 128)), []  # x axis (image sizes), y axis
+                for opt.imgsz in x:  # img-size
+                    LOGGER.info(f'\nRunning {f} --imgsz {opt.imgsz}...')
+                    r, _, t = run(**vars(opt), plots=False)
+                    y.append(r + t)  # results and times
+                np.savetxt(f, y, fmt='%10.4g')  # save
+            os.system('zip -r study.zip study_*.txt')
+            plot_val_study(x=x)  # plot
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

val_dual.py

@@ -0,0 +1,393 @@
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.callbacks import Callbacks
+from utils.dataloaders import create_dataloader
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, Profile, check_dataset, check_img_size, check_requirements,
+                           check_yaml, coco80_to_coco91_class, colorstr, increment_path, non_max_suppression,
+                           print_args, scale_boxes, xywh2xyxy, xyxy2xywh)
+from utils.metrics import ConfusionMatrix, ap_per_class, box_iou
+from utils.plots import output_to_target, plot_images, plot_val_study
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+def save_one_txt(predn, save_conf, shape, file):
+    # Save one txt result
+    gn = torch.tensor(shape)[[1, 0, 1, 0]]  # normalization gain whwh
+    for *xyxy, conf, cls in predn.tolist():
+        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
+        line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
+        with open(file, 'a') as f:
+            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+
+def save_one_json(predn, jdict, path, class_map):
+    # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}
+    image_id = int(path.stem) if path.stem.isnumeric() else path.stem
+    box = xyxy2xywh(predn[:, :4])  # xywh
+    box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
+    for p, b in zip(predn.tolist(), box.tolist()):
+        jdict.append({
+            'image_id': image_id,
+            'category_id': class_map[int(p[5])],
+            'bbox': [round(x, 3) for x in b],
+            'score': round(p[4], 5)})
+
+
+def process_batch(detections, labels, iouv):
+    """
+    Return correct prediction matrix
+    Arguments:
+        detections (array[N, 6]), x1, y1, x2, y2, conf, class
+        labels (array[M, 5]), class, x1, y1, x2, y2
+    Returns:
+        correct (array[N, 10]), for 10 IoU levels
+    """
+    correct = np.zeros((detections.shape[0], iouv.shape[0])).astype(bool)
+    iou = box_iou(labels[:, 1:], detections[:, :4])
+    correct_class = labels[:, 0:1] == detections[:, 5]
+    for i in range(len(iouv)):
+        x = torch.where((iou >= iouv[i]) & correct_class)  # IoU > threshold and classes match
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()  # [label, detect, iou]
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                # matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+            correct[matches[:, 1].astype(int), i] = True
+    return torch.tensor(correct, dtype=torch.bool, device=iouv.device)
+
+
+@smart_inference_mode()
+def run(
+        data,
+        weights=None,  # model.pt path(s)
+        batch_size=32,  # batch size
+        imgsz=640,  # inference size (pixels)
+        conf_thres=0.001,  # confidence threshold
+        iou_thres=0.7,  # NMS IoU threshold
+        max_det=300,  # maximum detections per image
+        task='val',  # train, val, test, speed or study
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        workers=8,  # max dataloader workers (per RANK in DDP mode)
+        single_cls=False,  # treat as single-class dataset
+        augment=False,  # augmented inference
+        verbose=False,  # verbose output
+        save_txt=False,  # save results to *.txt
+        save_hybrid=False,  # save label+prediction hybrid results to *.txt
+        save_conf=False,  # save confidences in --save-txt labels
+        save_json=False,  # save a COCO-JSON results file
+        project=ROOT / 'runs/val',  # save to project/name
+        name='exp',  # save to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        half=True,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        min_items=0,  # Experimental
+        model=None,
+        dataloader=None,
+        save_dir=Path(''),
+        plots=True,
+        callbacks=Callbacks(),
+        compute_loss=None,
+):
+    # Initialize/load model and set device
+    training = model is not None
+    if training:  # called by train.py
+        device, pt, jit, engine = next(model.parameters()).device, True, False, False  # get model device, PyTorch model
+        half &= device.type != 'cpu'  # half precision only supported on CUDA
+        model.half() if half else model.float()
+    else:  # called directly
+        device = select_device(device, batch_size=batch_size)
+
+        # Directories
+        save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+        (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+        # Load model
+        model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+        stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine
+        imgsz = check_img_size(imgsz, s=stride)  # check image size
+        half = model.fp16  # FP16 supported on limited backends with CUDA
+        if engine:
+            batch_size = model.batch_size
+        else:
+            device = model.device
+            if not (pt or jit):
+                batch_size = 1  # export.py models default to batch-size 1
+                LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models')
+
+        # Data
+        data = check_dataset(data)  # check
+
+    # Configure
+    model.eval()
+    cuda = device.type != 'cpu'
+    #is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'coco{os.sep}val2017.txt')  # COCO dataset
+    is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'val2017.txt')  # COCO dataset
+    nc = 1 if single_cls else int(data['nc'])  # number of classes
+    iouv = torch.linspace(0.5, 0.95, 10, device=device)  # iou vector for mAP@0.5:0.95
+    niou = iouv.numel()
+
+    # Dataloader
+    if not training:
+        if pt and not single_cls:  # check --weights are trained on --data
+            ncm = model.model.nc
+            assert ncm == nc, f'{weights} ({ncm} classes) trained on different --data than what you passed ({nc} ' \
+                              f'classes). Pass correct combination of --weights and --data that are trained together.'
+        model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz))  # warmup
+        pad, rect = (0.0, False) if task == 'speed' else (0.5, pt)  # square inference for benchmarks
+        task = task if task in ('train', 'val', 'test') else 'val'  # path to train/val/test images
+        dataloader = create_dataloader(data[task],
+                                       imgsz,
+                                       batch_size,
+                                       stride,
+                                       single_cls,
+                                       pad=pad,
+                                       rect=rect,
+                                       workers=workers,
+                                       min_items=opt.min_items,
+                                       prefix=colorstr(f'{task}: '))[0]
+
+    seen = 0
+    confusion_matrix = ConfusionMatrix(nc=nc)
+    names = model.names if hasattr(model, 'names') else model.module.names  # get class names
+    if isinstance(names, (list, tuple)):  # old format
+        names = dict(enumerate(names))
+    class_map = coco80_to_coco91_class() if is_coco else list(range(1000))
+    s = ('%22s' + '%11s' * 6) % ('Class', 'Images', 'Instances', 'P', 'R', 'mAP50', 'mAP50-95')
+    tp, fp, p, r, f1, mp, mr, map50, ap50, map = 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
+    dt = Profile(), Profile(), Profile()  # profiling times
+    loss = torch.zeros(3, device=device)
+    jdict, stats, ap, ap_class = [], [], [], []
+    callbacks.run('on_val_start')
+    pbar = tqdm(dataloader, desc=s, bar_format=TQDM_BAR_FORMAT)  # progress bar
+    for batch_i, (im, targets, paths, shapes) in enumerate(pbar):
+        callbacks.run('on_val_batch_start')
+        with dt[0]:
+            if cuda:
+                im = im.to(device, non_blocking=True)
+                targets = targets.to(device)
+            im = im.half() if half else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            nb, _, height, width = im.shape  # batch size, channels, height, width
+
+        # Inference
+        with dt[1]:
+            preds, train_out = model(im) if compute_loss else (model(im, augment=augment), None)
+
+        # Loss
+        if compute_loss:
+            preds = preds[1]
+            #train_out = train_out[1]
+            #loss += compute_loss(train_out, targets)[1]  # box, obj, cls
+        else:
+            preds = preds[0][1]
+
+        # NMS
+        targets[:, 2:] *= torch.tensor((width, height, width, height), device=device)  # to pixels
+        lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else []  # for autolabelling
+        with dt[2]:
+            preds = non_max_suppression(preds,
+                                        conf_thres,
+                                        iou_thres,
+                                        labels=lb,
+                                        multi_label=True,
+                                        agnostic=single_cls,
+                                        max_det=max_det)
+
+        # Metrics
+        for si, pred in enumerate(preds):
+            labels = targets[targets[:, 0] == si, 1:]
+            nl, npr = labels.shape[0], pred.shape[0]  # number of labels, predictions
+            path, shape = Path(paths[si]), shapes[si][0]
+            correct = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            seen += 1
+
+            if npr == 0:
+                if nl:
+                    stats.append((correct, *torch.zeros((2, 0), device=device), labels[:, 0]))
+                    if plots:
+                        confusion_matrix.process_batch(detections=None, labels=labels[:, 0])
+                continue
+
+            # Predictions
+            if single_cls:
+                pred[:, 5] = 0
+            predn = pred.clone()
+            scale_boxes(im[si].shape[1:], predn[:, :4], shape, shapes[si][1])  # native-space pred
+
+            # Evaluate
+            if nl:
+                tbox = xywh2xyxy(labels[:, 1:5])  # target boxes
+                scale_boxes(im[si].shape[1:], tbox, shape, shapes[si][1])  # native-space labels
+                labelsn = torch.cat((labels[:, 0:1], tbox), 1)  # native-space labels
+                correct = process_batch(predn, labelsn, iouv)
+                if plots:
+                    confusion_matrix.process_batch(predn, labelsn)
+            stats.append((correct, pred[:, 4], pred[:, 5], labels[:, 0]))  # (correct, conf, pcls, tcls)
+
+            # Save/log
+            if save_txt:
+                save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / f'{path.stem}.txt')
+            if save_json:
+                save_one_json(predn, jdict, path, class_map)  # append to COCO-JSON dictionary
+            callbacks.run('on_val_image_end', pred, predn, path, names, im[si])
+
+        # Plot images
+        if plots and batch_i < 3:
+            plot_images(im, targets, paths, save_dir / f'val_batch{batch_i}_labels.jpg', names)  # labels
+            plot_images(im, output_to_target(preds), paths, save_dir / f'val_batch{batch_i}_pred.jpg', names)  # pred
+
+        callbacks.run('on_val_batch_end', batch_i, im, targets, paths, shapes, preds)
+
+    # Compute metrics
+    stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)]  # to numpy
+    if len(stats) and stats[0].any():
+        tp, fp, p, r, f1, ap, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names)
+        ap50, ap = ap[:, 0], ap.mean(1)  # AP@0.5, AP@0.5:0.95
+        mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean()
+    nt = np.bincount(stats[3].astype(int), minlength=nc)  # number of targets per class
+
+    # Print results
+    pf = '%22s' + '%11i' * 2 + '%11.3g' * 4  # print format
+    LOGGER.info(pf % ('all', seen, nt.sum(), mp, mr, map50, map))
+    if nt.sum() == 0:
+        LOGGER.warning(f'WARNING ⚠️ no labels found in {task} set, can not compute metrics without labels')
+
+    # Print results per class
+    if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats):
+        for i, c in enumerate(ap_class):
+            LOGGER.info(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i]))
+
+    # Print speeds
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    if not training:
+        shape = (batch_size, 3, imgsz, imgsz)
+        LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t)
+
+    # Plots
+    if plots:
+        confusion_matrix.plot(save_dir=save_dir, names=list(names.values()))
+        callbacks.run('on_val_end', nt, tp, fp, p, r, f1, ap, ap50, ap_class, confusion_matrix)
+
+    # Save JSON
+    if save_json and len(jdict):
+        w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else ''  # weights
+        anno_json = str(Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json')  # annotations json
+        pred_json = str(save_dir / f"{w}_predictions.json")  # predictions json
+        LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...')
+        with open(pred_json, 'w') as f:
+            json.dump(jdict, f)
+
+        try:  # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
+            check_requirements('pycocotools')
+            from pycocotools.coco import COCO
+            from pycocotools.cocoeval import COCOeval
+
+            anno = COCO(anno_json)  # init annotations api
+            pred = anno.loadRes(pred_json)  # init predictions api
+            eval = COCOeval(anno, pred, 'bbox')
+            if is_coco:
+                eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files]  # image IDs to evaluate
+            eval.evaluate()
+            eval.accumulate()
+            eval.summarize()
+            map, map50 = eval.stats[:2]  # update results (mAP@0.5:0.95, mAP@0.5)
+        except Exception as e:
+            LOGGER.info(f'pycocotools unable to run: {e}')
+
+    # Return results
+    model.float()  # for training
+    if not training:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    maps = np.zeros(nc) + map
+    for i, c in enumerate(ap_class):
+        maps[c] = ap[i]
+    return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo.pt', help='model path(s)')
+    parser.add_argument('--batch-size', type=int, default=32, help='batch size')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.7, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=300, help='maximum detections per image')
+    parser.add_argument('--task', default='val', help='train, val, test, speed or study')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--verbose', action='store_true', help='report mAP by class')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-json', action='store_true', help='save a COCO-JSON results file')
+    parser.add_argument('--project', default=ROOT / 'runs/val', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    parser.add_argument('--min-items', type=int, default=0, help='Experimental')
+    opt = parser.parse_args()
+    opt.data = check_yaml(opt.data)  # check YAML
+    opt.save_json |= opt.data.endswith('coco.yaml')
+    opt.save_txt |= opt.save_hybrid
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    #check_requirements(exclude=('tensorboard', 'thop'))
+
+    if opt.task in ('train', 'val', 'test'):  # run normally
+        if opt.conf_thres > 0.001:  # https://github.com/ultralytics/yolov5/issues/1466
+            LOGGER.info(f'WARNING ⚠️ confidence threshold {opt.conf_thres} > 0.001 produces invalid results')
+        if opt.save_hybrid:
+            LOGGER.info('WARNING ⚠️ --save-hybrid will return high mAP from hybrid labels, not from predictions alone')
+        run(**vars(opt))
+
+    else:
+        weights = opt.weights if isinstance(opt.weights, list) else [opt.weights]
+        opt.half = torch.cuda.is_available() and opt.device != 'cpu'  # FP16 for fastest results
+        if opt.task == 'speed':  # speed benchmarks
+            # python val.py --task speed --data coco.yaml --batch 1 --weights yolo.pt...
+            opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False
+            for opt.weights in weights:
+                run(**vars(opt), plots=False)
+
+        elif opt.task == 'study':  # speed vs mAP benchmarks
+            # python val.py --task study --data coco.yaml --iou 0.7 --weights yolo.pt...
+            for opt.weights in weights:
+                f = f'study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt'  # filename to save to
+                x, y = list(range(256, 1536 + 128, 128)), []  # x axis (image sizes), y axis
+                for opt.imgsz in x:  # img-size
+                    LOGGER.info(f'\nRunning {f} --imgsz {opt.imgsz}...')
+                    r, _, t = run(**vars(opt), plots=False)
+                    y.append(r + t)  # results and times
+                np.savetxt(f, y, fmt='%10.4g')  # save
+            os.system('zip -r study.zip study_*.txt')
+            plot_val_study(x=x)  # plot
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)

val_triple.py

@@ -0,0 +1,391 @@
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+FILE = Path(__file__).resolve()
+ROOT = FILE.parents[0]  # YOLO root directory
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))  # add ROOT to PATH
+ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative
+
+from models.common import DetectMultiBackend
+from utils.callbacks import Callbacks
+from utils.dataloaders import create_dataloader
+from utils.general import (LOGGER, TQDM_BAR_FORMAT, Profile, check_dataset, check_img_size, check_requirements,
+                           check_yaml, coco80_to_coco91_class, colorstr, increment_path, non_max_suppression,
+                           print_args, scale_boxes, xywh2xyxy, xyxy2xywh)
+from utils.metrics import ConfusionMatrix, ap_per_class, box_iou
+from utils.plots import output_to_target, plot_images, plot_val_study
+from utils.torch_utils import select_device, smart_inference_mode
+
+
+def save_one_txt(predn, save_conf, shape, file):
+    # Save one txt result
+    gn = torch.tensor(shape)[[1, 0, 1, 0]]  # normalization gain whwh
+    for *xyxy, conf, cls in predn.tolist():
+        xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
+        line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
+        with open(file, 'a') as f:
+            f.write(('%g ' * len(line)).rstrip() % line + '\n')
+
+
+def save_one_json(predn, jdict, path, class_map):
+    # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}
+    image_id = int(path.stem) if path.stem.isnumeric() else path.stem
+    box = xyxy2xywh(predn[:, :4])  # xywh
+    box[:, :2] -= box[:, 2:] / 2  # xy center to top-left corner
+    for p, b in zip(predn.tolist(), box.tolist()):
+        jdict.append({
+            'image_id': image_id,
+            'category_id': class_map[int(p[5])],
+            'bbox': [round(x, 3) for x in b],
+            'score': round(p[4], 5)})
+
+
+def process_batch(detections, labels, iouv):
+    """
+    Return correct prediction matrix
+    Arguments:
+        detections (array[N, 6]), x1, y1, x2, y2, conf, class
+        labels (array[M, 5]), class, x1, y1, x2, y2
+    Returns:
+        correct (array[N, 10]), for 10 IoU levels
+    """
+    correct = np.zeros((detections.shape[0], iouv.shape[0])).astype(bool)
+    iou = box_iou(labels[:, 1:], detections[:, :4])
+    correct_class = labels[:, 0:1] == detections[:, 5]
+    for i in range(len(iouv)):
+        x = torch.where((iou >= iouv[i]) & correct_class)  # IoU > threshold and classes match
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()  # [label, detect, iou]
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                # matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+            correct[matches[:, 1].astype(int), i] = True
+    return torch.tensor(correct, dtype=torch.bool, device=iouv.device)
+
+
+@smart_inference_mode()
+def run(
+        data,
+        weights=None,  # model.pt path(s)
+        batch_size=32,  # batch size
+        imgsz=640,  # inference size (pixels)
+        conf_thres=0.001,  # confidence threshold
+        iou_thres=0.7,  # NMS IoU threshold
+        max_det=300,  # maximum detections per image
+        task='val',  # train, val, test, speed or study
+        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
+        workers=8,  # max dataloader workers (per RANK in DDP mode)
+        single_cls=False,  # treat as single-class dataset
+        augment=False,  # augmented inference
+        verbose=False,  # verbose output
+        save_txt=False,  # save results to *.txt
+        save_hybrid=False,  # save label+prediction hybrid results to *.txt
+        save_conf=False,  # save confidences in --save-txt labels
+        save_json=False,  # save a COCO-JSON results file
+        project=ROOT / 'runs/val',  # save to project/name
+        name='exp',  # save to project/name
+        exist_ok=False,  # existing project/name ok, do not increment
+        half=True,  # use FP16 half-precision inference
+        dnn=False,  # use OpenCV DNN for ONNX inference
+        min_items=0,  # Experimental
+        model=None,
+        dataloader=None,
+        save_dir=Path(''),
+        plots=True,
+        callbacks=Callbacks(),
+        compute_loss=None,
+):
+    # Initialize/load model and set device
+    training = model is not None
+    if training:  # called by train.py
+        device, pt, jit, engine = next(model.parameters()).device, True, False, False  # get model device, PyTorch model
+        half &= device.type != 'cpu'  # half precision only supported on CUDA
+        model.half() if half else model.float()
+    else:  # called directly
+        device = select_device(device, batch_size=batch_size)
+
+        # Directories
+        save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
+        (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir
+
+        # Load model
+        model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
+        stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine
+        imgsz = check_img_size(imgsz, s=stride)  # check image size
+        half = model.fp16  # FP16 supported on limited backends with CUDA
+        if engine:
+            batch_size = model.batch_size
+        else:
+            device = model.device
+            if not (pt or jit):
+                batch_size = 1  # export.py models default to batch-size 1
+                LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models')
+
+        # Data
+        data = check_dataset(data)  # check
+
+    # Configure
+    model.eval()
+    cuda = device.type != 'cpu'
+    #is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'coco{os.sep}val2017.txt')  # COCO dataset
+    is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'val2017.txt')  # COCO dataset
+    nc = 1 if single_cls else int(data['nc'])  # number of classes
+    iouv = torch.linspace(0.5, 0.95, 10, device=device)  # iou vector for mAP@0.5:0.95
+    niou = iouv.numel()
+
+    # Dataloader
+    if not training:
+        if pt and not single_cls:  # check --weights are trained on --data
+            ncm = model.model.nc
+            assert ncm == nc, f'{weights} ({ncm} classes) trained on different --data than what you passed ({nc} ' \
+                              f'classes). Pass correct combination of --weights and --data that are trained together.'
+        model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz))  # warmup
+        pad, rect = (0.0, False) if task == 'speed' else (0.5, pt)  # square inference for benchmarks
+        task = task if task in ('train', 'val', 'test') else 'val'  # path to train/val/test images
+        dataloader = create_dataloader(data[task],
+                                       imgsz,
+                                       batch_size,
+                                       stride,
+                                       single_cls,
+                                       pad=pad,
+                                       rect=rect,
+                                       workers=workers,
+                                       min_items=opt.min_items,
+                                       prefix=colorstr(f'{task}: '))[0]
+
+    seen = 0
+    confusion_matrix = ConfusionMatrix(nc=nc)
+    names = model.names if hasattr(model, 'names') else model.module.names  # get class names
+    if isinstance(names, (list, tuple)):  # old format
+        names = dict(enumerate(names))
+    class_map = coco80_to_coco91_class() if is_coco else list(range(1000))
+    s = ('%22s' + '%11s' * 6) % ('Class', 'Images', 'Instances', 'P', 'R', 'mAP50', 'mAP50-95')
+    tp, fp, p, r, f1, mp, mr, map50, ap50, map = 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
+    dt = Profile(), Profile(), Profile()  # profiling times
+    loss = torch.zeros(3, device=device)
+    jdict, stats, ap, ap_class = [], [], [], []
+    callbacks.run('on_val_start')
+    pbar = tqdm(dataloader, desc=s, bar_format=TQDM_BAR_FORMAT)  # progress bar
+    for batch_i, (im, targets, paths, shapes) in enumerate(pbar):
+        callbacks.run('on_val_batch_start')
+        with dt[0]:
+            if cuda:
+                im = im.to(device, non_blocking=True)
+                targets = targets.to(device)
+            im = im.half() if half else im.float()  # uint8 to fp16/32
+            im /= 255  # 0 - 255 to 0.0 - 1.0
+            nb, _, height, width = im.shape  # batch size, channels, height, width
+
+        # Inference
+        with dt[1]:
+            preds, train_out = model(im) if compute_loss else (model(im, augment=augment), None)
+            preds = preds[2]
+            train_out = train_out[2]
+
+        # Loss
+        #if compute_loss:
+        #    loss += compute_loss(train_out, targets)[2]  # box, obj, cls
+
+        # NMS
+        targets[:, 2:] *= torch.tensor((width, height, width, height), device=device)  # to pixels
+        lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else []  # for autolabelling
+        with dt[2]:
+            preds = non_max_suppression(preds,
+                                        conf_thres,
+                                        iou_thres,
+                                        labels=lb,
+                                        multi_label=True,
+                                        agnostic=single_cls,
+                                        max_det=max_det)
+
+        # Metrics
+        for si, pred in enumerate(preds):
+            labels = targets[targets[:, 0] == si, 1:]
+            nl, npr = labels.shape[0], pred.shape[0]  # number of labels, predictions
+            path, shape = Path(paths[si]), shapes[si][0]
+            correct = torch.zeros(npr, niou, dtype=torch.bool, device=device)  # init
+            seen += 1
+
+            if npr == 0:
+                if nl:
+                    stats.append((correct, *torch.zeros((2, 0), device=device), labels[:, 0]))
+                    if plots:
+                        confusion_matrix.process_batch(detections=None, labels=labels[:, 0])
+                continue
+
+            # Predictions
+            if single_cls:
+                pred[:, 5] = 0
+            predn = pred.clone()
+            scale_boxes(im[si].shape[1:], predn[:, :4], shape, shapes[si][1])  # native-space pred
+
+            # Evaluate
+            if nl:
+                tbox = xywh2xyxy(labels[:, 1:5])  # target boxes
+                scale_boxes(im[si].shape[1:], tbox, shape, shapes[si][1])  # native-space labels
+                labelsn = torch.cat((labels[:, 0:1], tbox), 1)  # native-space labels
+                correct = process_batch(predn, labelsn, iouv)
+                if plots:
+                    confusion_matrix.process_batch(predn, labelsn)
+            stats.append((correct, pred[:, 4], pred[:, 5], labels[:, 0]))  # (correct, conf, pcls, tcls)
+
+            # Save/log
+            if save_txt:
+                save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / f'{path.stem}.txt')
+            if save_json:
+                save_one_json(predn, jdict, path, class_map)  # append to COCO-JSON dictionary
+            callbacks.run('on_val_image_end', pred, predn, path, names, im[si])
+
+        # Plot images
+        if plots and batch_i < 3:
+            plot_images(im, targets, paths, save_dir / f'val_batch{batch_i}_labels.jpg', names)  # labels
+            plot_images(im, output_to_target(preds), paths, save_dir / f'val_batch{batch_i}_pred.jpg', names)  # pred
+
+        callbacks.run('on_val_batch_end', batch_i, im, targets, paths, shapes, preds)
+
+    # Compute metrics
+    stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)]  # to numpy
+    if len(stats) and stats[0].any():
+        tp, fp, p, r, f1, ap, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names)
+        ap50, ap = ap[:, 0], ap.mean(1)  # AP@0.5, AP@0.5:0.95
+        mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean()
+    nt = np.bincount(stats[3].astype(int), minlength=nc)  # number of targets per class
+
+    # Print results
+    pf = '%22s' + '%11i' * 2 + '%11.3g' * 4  # print format
+    LOGGER.info(pf % ('all', seen, nt.sum(), mp, mr, map50, map))
+    if nt.sum() == 0:
+        LOGGER.warning(f'WARNING ⚠️ no labels found in {task} set, can not compute metrics without labels')
+
+    # Print results per class
+    if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats):
+        for i, c in enumerate(ap_class):
+            LOGGER.info(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i]))
+
+    # Print speeds
+    t = tuple(x.t / seen * 1E3 for x in dt)  # speeds per image
+    if not training:
+        shape = (batch_size, 3, imgsz, imgsz)
+        LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t)
+
+    # Plots
+    if plots:
+        confusion_matrix.plot(save_dir=save_dir, names=list(names.values()))
+        callbacks.run('on_val_end', nt, tp, fp, p, r, f1, ap, ap50, ap_class, confusion_matrix)
+
+    # Save JSON
+    if save_json and len(jdict):
+        w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else ''  # weights
+        anno_json = str(Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json')  # annotations json
+        pred_json = str(save_dir / f"{w}_predictions.json")  # predictions json
+        LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...')
+        with open(pred_json, 'w') as f:
+            json.dump(jdict, f)
+
+        try:  # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
+            check_requirements('pycocotools')
+            from pycocotools.coco import COCO
+            from pycocotools.cocoeval import COCOeval
+
+            anno = COCO(anno_json)  # init annotations api
+            pred = anno.loadRes(pred_json)  # init predictions api
+            eval = COCOeval(anno, pred, 'bbox')
+            if is_coco:
+                eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files]  # image IDs to evaluate
+            eval.evaluate()
+            eval.accumulate()
+            eval.summarize()
+            map, map50 = eval.stats[:2]  # update results (mAP@0.5:0.95, mAP@0.5)
+        except Exception as e:
+            LOGGER.info(f'pycocotools unable to run: {e}')
+
+    # Return results
+    model.float()  # for training
+    if not training:
+        s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
+        LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
+    maps = np.zeros(nc) + map
+    for i, c in enumerate(ap_class):
+        maps[c] = ap[i]
+    return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t
+
+
+def parse_opt():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data', type=str, default=ROOT / 'data/coco.yaml', help='dataset.yaml path')
+    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolo.pt', help='model path(s)')
+    parser.add_argument('--batch-size', type=int, default=32, help='batch size')
+    parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold')
+    parser.add_argument('--iou-thres', type=float, default=0.7, help='NMS IoU threshold')
+    parser.add_argument('--max-det', type=int, default=300, help='maximum detections per image')
+    parser.add_argument('--task', default='val', help='train, val, test, speed or study')
+    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
+    parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)')
+    parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset')
+    parser.add_argument('--augment', action='store_true', help='augmented inference')
+    parser.add_argument('--verbose', action='store_true', help='report mAP by class')
+    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
+    parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt')
+    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
+    parser.add_argument('--save-json', action='store_true', help='save a COCO-JSON results file')
+    parser.add_argument('--project', default=ROOT / 'runs/val', help='save to project/name')
+    parser.add_argument('--name', default='exp', help='save to project/name')
+    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
+    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
+    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
+    parser.add_argument('--min-items', type=int, default=0, help='Experimental')
+    opt = parser.parse_args()
+    opt.data = check_yaml(opt.data)  # check YAML
+    opt.save_json |= opt.data.endswith('coco.yaml')
+    opt.save_txt |= opt.save_hybrid
+    print_args(vars(opt))
+    return opt
+
+
+def main(opt):
+    #check_requirements(exclude=('tensorboard', 'thop'))
+
+    if opt.task in ('train', 'val', 'test'):  # run normally
+        if opt.conf_thres > 0.001:  # https://github.com/ultralytics/yolov5/issues/1466
+            LOGGER.info(f'WARNING ⚠️ confidence threshold {opt.conf_thres} > 0.001 produces invalid results')
+        if opt.save_hybrid:
+            LOGGER.info('WARNING ⚠️ --save-hybrid will return high mAP from hybrid labels, not from predictions alone')
+        run(**vars(opt))
+
+    else:
+        weights = opt.weights if isinstance(opt.weights, list) else [opt.weights]
+        opt.half = torch.cuda.is_available() and opt.device != 'cpu'  # FP16 for fastest results
+        if opt.task == 'speed':  # speed benchmarks
+            # python val.py --task speed --data coco.yaml --batch 1 --weights yolo.pt...
+            opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False
+            for opt.weights in weights:
+                run(**vars(opt), plots=False)
+
+        elif opt.task == 'study':  # speed vs mAP benchmarks
+            # python val.py --task study --data coco.yaml --iou 0.7 --weights yolo.pt...
+            for opt.weights in weights:
+                f = f'study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt'  # filename to save to
+                x, y = list(range(256, 1536 + 128, 128)), []  # x axis (image sizes), y axis
+                for opt.imgsz in x:  # img-size
+                    LOGGER.info(f'\nRunning {f} --imgsz {opt.imgsz}...')
+                    r, _, t = run(**vars(opt), plots=False)
+                    y.append(r + t)  # results and times
+                np.savetxt(f, y, fmt='%10.4g')  # save
+            os.system('zip -r study.zip study_*.txt')
+            plot_val_study(x=x)  # plot
+
+
+if __name__ == "__main__":
+    opt = parse_opt()
+    main(opt)