File size: 12,480 Bytes
7f51798
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
"""
Train a diffusion model on images.
"""
from pdb import set_trace as st
import random
import json
import sys
import os

sys.path.append('.')

import torch.distributed as dist

import traceback

import torch as th

# if th.cuda.is_available(): # FIXME
#     from xformers.triton import FusedLayerNorm as LayerNorm

import torch.multiprocessing as mp
import numpy as np

import argparse
import dnnlib
from guided_diffusion import dist_util, logger
from guided_diffusion.script_util import (
    args_to_dict,
    add_dict_to_argparser,
)
from nsr.train_nv_util import TrainLoop3DRecNV, TrainLoop3DRec, TrainLoop3DRecNVPatch, TrainLoop3DRecNVPatchSingleForward, TrainLoop3DRecNVPatchSingleForwardMV, TrainLoop3DRecNVPatchSingleForwardMV_NoCrop, TrainLoop3DRecNVPatchSingleForwardMVAdvLoss, TrainLoop3DRecNVPatchSingleForwardMV_NoCrop_adv
from nsr.script_util import create_3DAE_model, encoder_and_nsr_defaults, loss_defaults, rendering_options_defaults, eg3d_options_default, dataset_defaults
# from datasets import g_buffer_objaverse
from nsr.losses.builder import E3DGELossClass, E3DGE_with_AdvLoss


# th.backends.cuda.matmul.allow_tf32 = True # https://huggingface.co/docs/diffusers/optimization/fp16
th.backends.cuda.matmul.allow_tf32 = True
th.backends.cudnn.allow_tf32 = True
th.backends.cudnn.enabled = True


def training_loop(args):
    # def training_loop(args):
    dist_util.setup_dist(args)
    # th.autograd.set_detect_anomaly(True) # type: ignore
    th.autograd.set_detect_anomaly(False)  # type: ignore
    # https://blog.csdn.net/qq_41682740/article/details/126304613

    SEED = args.seed

    # dist.init_process_group(backend='nccl', init_method='env://', rank=args.local_rank, world_size=th.cuda.device_count())
    logger.log(f"global_rank={args.global_rank}, local_rank={args.local_rank} init complete, seed={SEED}")
    th.cuda.set_device(args.local_rank)
    th.cuda.empty_cache()

    # * deterministic algorithms flags
    th.cuda.manual_seed_all(SEED)
    np.random.seed(SEED)
    random.seed(SEED)

    # logger.configure(dir=args.logdir, format_strs=["tensorboard", "csv"])
    logger.configure(dir=args.logdir)

    logger.log("creating encoder and NSR decoder...")
    # device = dist_util.dev()
    device = th.device("cuda", args.local_rank)

    # shared eg3d opts
    opts = eg3d_options_default()

    if args.sr_training:
        args.sr_kwargs = dnnlib.EasyDict(
            channel_base=opts.cbase,
            channel_max=opts.cmax,
            fused_modconv_default='inference_only',
            use_noise=True
        )  # ! close noise injection? since noise_mode='none' in eg3d

    auto_encoder = create_3DAE_model(
        **args_to_dict(args,
                       encoder_and_nsr_defaults().keys()))
    auto_encoder.to(device)
    auto_encoder.train()

    logger.log("creating data loader...")
    # data = load_data(
    # st()
    if args.objv_dataset:
        from datasets.g_buffer_objaverse import load_data, load_eval_data, load_memory_data, load_wds_data
    else:  # shapenet
        from datasets.shapenet import load_data, load_eval_data, load_memory_data

    if args.overfitting:
        data = load_memory_data(
            file_path=args.data_dir,
            batch_size=args.batch_size,
            reso=args.image_size,
            reso_encoder=args.image_size_encoder,  # 224 -> 128
            num_workers=args.num_workers,
            # load_depth=args.depth_lambda > 0
            # load_depth=True,  # for evaluation
                    **args_to_dict(args,
                                   dataset_defaults().keys()))
        eval_data = None
    else:
        if args.use_wds:
            # st()
            if args.data_dir == 'NONE':
                with open(args.shards_lst) as f:
                    shards_lst = [url.strip() for url in f.readlines()]
                data = load_wds_data(
                    shards_lst,  # type: ignore
                    args.image_size,
                    args.image_size_encoder,
                    args.batch_size,
                    args.num_workers,
                    #  plucker_embedding=args.plucker_embedding,
                    #  mv_input=args.mv_input,
                    #  split_chunk_input=args.split_chunk_input,
                    **args_to_dict(args,
                                   dataset_defaults().keys()))

            elif not args.inference:
                data = load_wds_data(args.data_dir,
                                     args.image_size,
                                     args.image_size_encoder,
                                     args.batch_size,
                                     args.num_workers,
                                     plucker_embedding=args.plucker_embedding,
                                     mv_input=args.mv_input,
                                     split_chunk_input=args.split_chunk_input)
            else:
                data = None
            # ! load eval

            if args.eval_data_dir == 'NONE':
                with open(args.eval_shards_lst) as f:
                    eval_shards_lst = [url.strip() for url in f.readlines()]
            else:
                eval_shards_lst = args.eval_data_dir  # auto expanded

            eval_data = load_wds_data(
                eval_shards_lst,  # type: ignore
                args.image_size,
                args.image_size_encoder,
                args.eval_batch_size,
                args.num_workers,
                **args_to_dict(args,
                               dataset_defaults().keys()))
            # load_instance=True) # TODO

        else:

            if args.inference:
                data = None
            else:
                data = load_data(
                    file_path=args.data_dir,
                    batch_size=args.batch_size,
                    reso=args.image_size,
                    reso_encoder=args.image_size_encoder,  # 224 -> 128
                    num_workers=args.num_workers,
                    **args_to_dict(args,
                                   dataset_defaults().keys())
                                   )

                    # load_depth=True # for evaluation

            if args.pose_warm_up_iter > 0:
                overfitting_dataset = load_memory_data(
                    file_path=args.data_dir,
                    batch_size=args.batch_size,
                    reso=args.image_size,
                    reso_encoder=args.image_size_encoder,  # 224 -> 128
                    num_workers=args.num_workers,
                    # load_depth=args.depth_lambda > 0
                    # load_depth=True  # for evaluation
                    **args_to_dict(args,
                                   dataset_defaults().keys()))
                data = [data, overfitting_dataset, args.pose_warm_up_iter]

            eval_data = load_eval_data(
                file_path=args.eval_data_dir,
                batch_size=args.eval_batch_size,
                reso=args.image_size,
                reso_encoder=args.image_size_encoder,  # 224 -> 128
                num_workers=args.num_workers,
                load_depth=True,  # for evaluation
                preprocess=auto_encoder.preprocess,
                wds_split=args.wds_split,
                # interval=args.interval,
                # use_lmdb=args.use_lmdb,
                # plucker_embedding=args.plucker_embedding,
                # load_real=args.load_real,
                # four_view_for_latent=args.four_view_for_latent,
                # load_extra_36_view=args.load_extra_36_view,
                # shuffle_across_cls=args.shuffle_across_cls,
                **args_to_dict(args,
                               dataset_defaults().keys()))

    logger.log("creating data loader done...")

    args.img_size = [args.image_size_encoder]
    # try dry run
    # batch = next(data)
    # batch = None

    # logger.log("creating model and diffusion...")

    # let all processes sync up before starting with a new epoch of training
    dist_util.synchronize()

    # schedule_sampler = create_named_schedule_sampler(args.schedule_sampler, diffusion)

    opt = dnnlib.EasyDict(args_to_dict(args, loss_defaults().keys()))
    # opt.max_depth, opt.min_depth = args.rendering_kwargs.ray_end, args.rendering_kwargs.ray_start
    if 'disc' in args.trainer_name:
        loss_class = E3DGE_with_AdvLoss(
            device,
            opt,
            # disc_weight=args.patchgan_disc, # rec_cvD_lambda
            disc_factor=args.patchgan_disc_factor,  # reduce D update speed
            disc_weight=args.patchgan_disc_g_weight).to(device)
    else:
        loss_class = E3DGELossClass(device, opt).to(device)

    # writer = SummaryWriter() # TODO, add log dir

    logger.log("training...")

    TrainLoop = {
        'input_rec': TrainLoop3DRec,
        'nv_rec': TrainLoop3DRecNV,
        # 'nv_rec_patch': TrainLoop3DRecNVPatch,
        'nv_rec_patch': TrainLoop3DRecNVPatchSingleForward,
        'nv_rec_patch_mvE': TrainLoop3DRecNVPatchSingleForwardMV,
        'nv_rec_patch_mvE_disc': TrainLoop3DRecNVPatchSingleForwardMVAdvLoss,
        'nv_rec_patch_mvE_gs': TrainLoop3DRecNVPatchSingleForwardMV_NoCrop,
        'nv_rec_patch_mvE_gs_disc': TrainLoop3DRecNVPatchSingleForwardMV_NoCrop_adv,
    }[args.trainer_name]

    logger.log("creating TrainLoop done...")

    # th._dynamo.config.verbose=True # th212 required
    # th._dynamo.config.suppress_errors = True
    auto_encoder.decoder.rendering_kwargs = args.rendering_kwargs
    train_loop = TrainLoop(
        rec_model=auto_encoder,
        loss_class=loss_class,
        data=data,
        eval_data=eval_data,
        #   compile=args.compile,
        **vars(args))
    # train_loop.rendering_kwargs = args.rendering_kwargs

    if args.inference:
        camera = th.load('eval_pose.pt', map_location=dist_util.dev())
        train_loop.eval_novelview_loop(camera=camera,
                                       save_latent=args.save_latent)
    else:
        train_loop.run_loop()


def create_argparser(**kwargs):
    # defaults.update(model_and_diffusion_defaults())

    defaults = dict(
        seed=0,
        dataset_size=-1,
        trainer_name='input_rec',
        use_amp=False,
        overfitting=False,
        num_workers=4,
        image_size=128,
        image_size_encoder=224,
        iterations=150000,
        anneal_lr=False,
        lr=5e-5,
        weight_decay=0.0,
        lr_anneal_steps=0,
        batch_size=1,
        eval_batch_size=12,
        microbatch=-1,  # -1 disables microbatches
        ema_rate="0.9999",  # comma-separated list of EMA values
        log_interval=50,
        eval_interval=2500,
        save_interval=10000,
        resume_checkpoint="",
        use_fp16=False,
        fp16_scale_growth=1e-3,
        data_dir="",
        eval_data_dir="",
        # load_depth=False, # TODO
        logdir="/mnt/lustre/yslan/logs/nips23/",
        # test warm up pose sampling training
        pose_warm_up_iter=-1,
        inference=False,
        export_latent=False,
        save_latent=False,
        wds_split=1,  # out of 4
    )

    defaults.update(dataset_defaults())  # type: ignore
    defaults.update(encoder_and_nsr_defaults())  # type: ignore
    defaults.update(loss_defaults())

    parser = argparse.ArgumentParser()
    add_dict_to_argparser(parser, defaults)

    return parser


if __name__ == "__main__":
    th.multiprocessing.set_start_method('spawn')

    # os.environ[
    # "TORCH_DISTRIBUTED_DEBUG"] = "DETAIL"  # set to DETAIL for runtime logging.
    # os.environ["TORCH_CPP_LOG_LEVEL"]="INFO"
    # os.environ["NCCL_DEBUG"]="INFO"

    args = create_argparser().parse_args()
    args.local_rank = int(os.environ["LOCAL_RANK"])
    # if os.environ['WORLD_SIZE'] > 1:

    # for multi-node training
    if dist_util.get_world_size() > 1:
        args.global_rank = int(os.environ["RANK"])
    else:
        args.global_rank = 0

    args.gpus = th.cuda.device_count()

    opts = args

    args.rendering_kwargs = rendering_options_defaults(opts)

    # print(args)
    with open(os.path.join(args.logdir, 'args.json'), 'w') as f:
        json.dump(vars(args), f, indent=2)

    # Launch processes.
    print('Launching processes...')

    try:
        training_loop(args)
    # except KeyboardInterrupt as e:
    except Exception as e:
        # print(e)
        traceback.print_exc()
        dist_util.cleanup()  # clean port and socket when ctrl+c