Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,727 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 |
"""
Train a diffusion model on images.
"""
import sys
import os
sys.path.append('.')
import torch.distributed as dist
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, loss_defaults
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
th.backends.cuda.matmul.allow_tf32 = True
th.backends.cudnn.allow_tf32 = True
th.backends.cudnn.enabled = True
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.manual_seed_all(SEED)
# logger.configure(dir=args.logdir, format_strs=["tensorboard", "csv"])
logger.configure(dir=args.logdir)
logger.log("creating data loader...")
# TODO, load shapenet data
# 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,
# load_depth=args.depth_lambda > 0
# # load_depth=True # for evaluation
# )
# 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,
# load_depth=args.depth_lambda > 0
# # load_depth=True # for evaluation
# )
eval_data = load_eval_data(
file_path=args.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
)
# try dry run
# batch = next(data)
# batch = None
# logger.log("creating model and diffusion...")
logger.log("creating encoder and NSR decoder...")
# device = dist_util.dev()
device = th.device("cuda", args.local_rank)
auto_encoder = create_3DAE_model(
**args_to_dict(args,
encoder_and_nsr_defaults().keys()))
auto_encoder.to(device)
auto_encoder.eval()
# dist_util.sync_params(auto_encoder.named_parameters())
# auto_encoder.train()
# let all processes sync up before starting with a new epoch of training
dist_util.synchronize()
# noise = th.randn(1, 14 * 14, 384).to(device) # B, L, C
# noise = th.randn(1, 3,224,224).to(device)
# img = auto_encoder(noise, th.zeros(1, 25).to(device))
# print(img['image'].shape)
# if dist_util.get_rank()==0:
# print(auto_encoder)
# 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)
# logger.log("training...")
TrainLoop(
rec_model=auto_encoder,
loss_class=loss_class,
# diffusion=diffusion,
data=None,
eval_interval=-1,
eval_data=eval_data,
# 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,
resume_cldm_checkpoint=args.resume_cldm_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(
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=8,
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/",
)
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()
args = create_argparser().parse_args()
args.local_rank = int(os.environ["LOCAL_RANK"])
args.gpus = th.cuda.device_count()
args.master_addr = master_addr
args.master_port = master_port
# Launch processes.
print('Launching processes...')
training_loop(args)
|