File size: 4,505 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
"""
Train a diffusion model on images.
"""
import sys
import os
sys.path.append('.')

import torch as th
import torch.multiprocessing as mp

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
from nsr.script_util import create_3DAE_model, encoder_and_nsr_defaults, create_Triplane, loss_defaults
from datasets.shapenet import load_eval_rays, load_data, load_eval_data
from nsr.losses.builder import E3DGELossClass

from pdb import set_trace as st

def inference_loop(rank, master_addr, master_port, args):
    dist_util.setup_dist(rank, master_addr, master_port, args.gpus)

    logger.configure(dir=args.logdir)

    logger.log("creating eval rays...")
    # TODO, load shapenet data
    eval_data = load_eval_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
    )
    # c_list = load_eval_rays(
    #     file_path=args.data_dir,
    #     reso=args.image_size,
    #     reso_encoder=args.image_size_encoder, # 224 -> 128
    # )

    # try dry run
    # batch = next(data)
    # batch = None

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

    logger.log("loading encoder and NSR decoder...") 
    auto_encoder = create_Triplane( # basically overfitting tirplane
        **args_to_dict(args,
                       encoder_and_nsr_defaults().keys()))

    # auto_encoder = create_3DAE_model(
    #     **args_to_dict(args,
    #                    encoder_and_nsr_defaults().keys()))
    auto_encoder.to(dist_util.dev())
    auto_encoder.eval()

    # schedule_sampler = create_named_schedule_sampler(args.schedule_sampler, diffusion)


    opt = dnnlib.EasyDict(args_to_dict(args, loss_defaults().keys()) )
    loss_class = E3DGELossClass(dist_util.dev(), opt).to(dist_util.dev())

    logger.log("training...")
    TrainLoop(
        model=auto_encoder,
        # encoder,
        # decoder
        loss_class=loss_class,
        # diffusion=diffusion,
        data=eval_data, # TODO
        # data=batch,
        batch_size=args.batch_size,
        microbatch=args.microbatch,
        lr=args.lr,
        ema_rate=args.ema_rate,
        log_interval=args.log_interval,
        save_interval=args.save_interval,
        resume_checkpoint=args.resume_checkpoint,
        use_fp16=args.use_fp16,
        fp16_scale_growth=args.fp16_scale_growth,
        weight_decay=args.weight_decay,
        lr_anneal_steps=args.lr_anneal_steps,
    ).eval_loop() # ! overfitting


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

    defaults = dict(
        num_workers=4,
        local_rank=0,
        gpus=1,
        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,
        microbatch=-1,  # -1 disables microbatches
        ema_rate="0.9999",  # comma-separated list of EMA values
        log_interval=10,
        save_interval=10000,
        resume_checkpoint="",
        use_fp16=False,
        fp16_scale_growth=1e-3,
        data_dir="",
        # load_depth=False, # TODO
        logdir="/mnt/lustre/yslan/logs/nips23/eval",
    )

    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"

    args = create_argparser().parse_args()
    # st()

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

    # Launch processes.
    print('Launching processes...')
    th.multiprocessing.set_start_method('spawn')

    subprocess_fn = inference_loop

    # launch using torch.multiprocessing.spawn
    if args.gpus == 1:
        subprocess_fn(rank=0, master_addr=master_addr, master_port=master_port, args=args)
    else:
        th.multiprocessing.spawn(fn=subprocess_fn,
                                    args=(master_addr, master_port,args),
                                    nprocs=args.gpus)