File size: 3,159 Bytes
88359db |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 |
import time
from urllib.request import urlopen
import cupy as cp
import numpy as np
import onnxruntime as ort
from PIL import Image
img = Image.open(
urlopen(
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png"
)
)
def transforms_numpy(image: Image.Image):
image = image.convert("RGB")
image = image.resize((448, 448), Image.BICUBIC)
img_numpy = np.array(image).astype(np.float32) / 255.0
img_numpy = img_numpy.transpose(2, 0, 1)
mean = np.array([0.4815, 0.4578, 0.4082]).reshape(-1, 1, 1)
std = np.array([0.2686, 0.2613, 0.2758]).reshape(-1, 1, 1)
img_numpy = (img_numpy - mean) / std
img_numpy = np.expand_dims(img_numpy, axis=0)
img_numpy = img_numpy.astype(np.float32)
return img_numpy
def transforms_cupy(image: Image.Image):
# Convert image to RGB and resize
image = image.convert("RGB")
image = image.resize((448, 448), Image.BICUBIC)
# Convert to CuPy array and normalize
img_cupy = cp.array(image, dtype=cp.float32) / 255.0
img_cupy = img_cupy.transpose(2, 0, 1)
# Apply mean and std normalization
mean = cp.array([0.4815, 0.4578, 0.4082], dtype=cp.float32).reshape(-1, 1, 1)
std = cp.array([0.2686, 0.2613, 0.2758], dtype=cp.float32).reshape(-1, 1, 1)
img_cupy = (img_cupy - mean) / std
# Add batch dimension
img_cupy = cp.expand_dims(img_cupy, axis=0)
return img_cupy
# Create ONNX Runtime session with CPU provider
onnx_filename = "eva02_large_patch14_448.onnx"
providers = [
(
"TensorrtExecutionProvider",
{
"device_id": 0,
"trt_max_workspace_size": 8589934592,
"trt_fp16_enable": True,
"trt_engine_cache_enable": True,
"trt_engine_cache_path": "./trt_cache",
"trt_force_sequential_engine_build": False,
"trt_max_partition_iterations": 10000,
"trt_min_subgraph_size": 1,
"trt_builder_optimization_level": 5,
"trt_timing_cache_enable": True,
},
),
]
session = ort.InferenceSession(onnx_filename, providers=providers)
# Get input and output names
input_name = session.get_inputs()[0].name
output_name = session.get_outputs()[0].name
# Run inference
output = session.run([output_name], {input_name: transforms_numpy(img)})[0]
# Run benchmark numpy
num_images = 100
start = time.perf_counter()
for i in range(num_images):
output = session.run([output_name], {input_name: transforms_numpy(img)})[0]
end = time.perf_counter()
time_taken = end - start
ms_per_image = time_taken / num_images * 1000
fps = num_images / time_taken
print(f"TensorRT + numpy: {ms_per_image:.3f} ms per image, FPS: {fps:.2f}")
# Run benchmark cupy
num_images = 100
start = time.perf_counter()
for i in range(num_images):
img_cupy = transforms_cupy(img)
output = session.run([output_name], {input_name: cp.asnumpy(img_cupy)})[0]
end = time.perf_counter()
time_taken = end - start
ms_per_image = time_taken / num_images * 1000
fps = num_images / time_taken
print(f"TensorRT + cupy : {ms_per_image:.3f} ms per image, FPS: {fps:.2f}")
|