File size: 6,693 Bytes
87c126b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Train a diffusion model on images.
"""
import json
import sys
import os

sys.path.append('.')
import torch.distributed as dist

import traceback

import torch as th
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_util import TrainLoop3DRec as TrainLoop

import nsr
from nsr.script_util import create_3DAE_model, encoder_and_nsr_defaults, loss_defaults, rendering_options_defaults, eg3d_options_default
from datasets.shapenet import load_data, load_eval_data, load_memory_data
from nsr.losses.builder import E3DGELossClass

from pdb import set_trace as st

import warnings

warnings.filterwarnings("ignore", category=UserWarning)

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

SEED = 0


def training_loop(args):
    # def training_loop(args):
    dist_util.setup_dist(args)

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

    th.cuda.manual_seed_all(SEED)
    np.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(
    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,
            # trainer_name=args.trainer_name,
            # load_depth=args.depth_lambda > 0
            load_depth=True  # for evaluation
        )
    else:
        data = load_data(
            dataset_size=args.dataset_size,
            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=True,
            preprocess=auto_encoder.preprocess,  # clip
            trainer_name=args.trainer_name,
            use_lmdb=args.use_lmdb
            # load_depth=True # for evaluation
        )
    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=2,
        load_depth=True,  # for evaluation
        preprocess=auto_encoder.preprocess)
    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()))
    loss_class = E3DGELossClass(device, opt).to(device)

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

    logger.log("training...")

    TrainLoop = {
        'cvD': nsr.TrainLoop3DcvD,
        'nvsD': nsr.TrainLoop3DcvD_nvsD,
        'nvsD_nosr': nsr.TrainLoop3DcvD_nvsD_noSR,
        'cano_nvsD_nosr': nsr.TrainLoop3DcvD_nvsD_noSR,
        'cano_nvs_cvD': nsr.TrainLoop3DcvD_nvsD_canoD,
        'cano_nvs_cvD_nv': nsr.TrainLoop3DcvD_nvsD_canoD_multiview,
        'cvD_nvsD_canoD_canomask': nsr.TrainLoop3DcvD_nvsD_canoD_canomask,
        'canoD': nsr.TrainLoop3DcvD_canoD
    }[args.trainer_name]

    TrainLoop(rec_model=auto_encoder,
              loss_class=loss_class,
              data=data,
              eval_data=eval_data,
              **vars(args)).run_loop()  # ! overfitting


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

    defaults = dict(
        dataset_size=-1,
        trainer_name='cvD',
        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/",
        pose_warm_up_iter=-1,
        use_lmdb=False,
    )

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

    # master_addr = '127.0.0.1'
    # master_port = dist_util._find_free_port()
    # master_port = 31323

    args = create_argparser().parse_args()
    args.local_rank = int(os.environ["LOCAL_RANK"])
    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