/
train_ldm.py
165 lines (136 loc) · 5.56 KB
/
train_ldm.py
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
""" Training script for the diffusion model in the latent space of the pretraine AEKL model. """
import argparse
import warnings
from pathlib import Path
import mlflow.pytorch
import torch
import torch.nn as nn
import torch.optim as optim
from generative.networks.nets import DiffusionModelUNet
from generative.networks.schedulers import DDPMScheduler
from monai.config import print_config
from monai.utils import set_determinism
from omegaconf import OmegaConf
from tensorboardX import SummaryWriter
from training_functions import train_ldm
from transformers import CLIPTextModel
from util import get_dataloader, log_mlflow
warnings.filterwarnings("ignore")
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument("--seed", type=int, default=2, help="Random seed to use.")
parser.add_argument("--run_dir", help="Location of model to resume.")
parser.add_argument("--training_ids", help="Location of file with training ids.")
parser.add_argument("--validation_ids", help="Location of file with validation ids.")
parser.add_argument("--config_file", help="Location of file with validation ids.")
parser.add_argument("--stage1_uri", help="Path readable by load_model.")
parser.add_argument("--scale_factor", type=float, help="Path readable by load_model.")
parser.add_argument("--batch_size", type=int, default=256, help="Training batch size.")
parser.add_argument("--n_epochs", type=int, default=25, help="Number of epochs to train.")
parser.add_argument("--eval_freq", type=int, default=10, help="Number of epochs to between evaluations.")
parser.add_argument("--num_workers", type=int, default=8, help="Number of loader workers")
parser.add_argument("--experiment", help="Mlflow experiment name.")
args = parser.parse_args()
return args
class Stage1Wrapper(nn.Module):
"""Wrapper for stage 1 model as a workaround for the DataParallel usage in the training loop."""
def __init__(self, model: nn.Module) -> None:
super().__init__()
self.model = model
def forward(self, x: torch.Tensor) -> torch.Tensor:
z_mu, z_sigma = self.model.encode(x)
z = self.model.sampling(z_mu, z_sigma)
return z
def main(args):
set_determinism(seed=args.seed)
print_config()
output_dir = Path("/project/outputs/runs/")
output_dir.mkdir(exist_ok=True, parents=True)
run_dir = output_dir / args.run_dir
if run_dir.exists() and (run_dir / "checkpoint.pth").exists():
resume = True
else:
resume = False
run_dir.mkdir(exist_ok=True)
print(f"Run directory: {str(run_dir)}")
print(f"Arguments: {str(args)}")
for k, v in vars(args).items():
print(f" {k}: {v}")
writer_train = SummaryWriter(log_dir=str(run_dir / "train"))
writer_val = SummaryWriter(log_dir=str(run_dir / "val"))
print("Getting data...")
cache_dir = output_dir / "cached_data_diffusion"
cache_dir.mkdir(exist_ok=True)
train_loader, val_loader = get_dataloader(
cache_dir=cache_dir,
batch_size=args.batch_size,
training_ids=args.training_ids,
validation_ids=args.validation_ids,
num_workers=args.num_workers,
model_type="diffusion",
)
# Load Autoencoder to produce the latent representations
print(f"Loading Stage 1 from {args.stage1_uri}")
stage1 = mlflow.pytorch.load_model(args.stage1_uri)
stage1 = Stage1Wrapper(model=stage1)
stage1.eval()
# Create the diffusion model
print("Creating model...")
config = OmegaConf.load(args.config_file)
diffusion = DiffusionModelUNet(**config["ldm"].get("params", dict()))
scheduler = DDPMScheduler(**config["ldm"].get("scheduler", dict()))
text_encoder = CLIPTextModel.from_pretrained("stabilityai/stable-diffusion-2-1-base", subfolder="text_encoder")
print(f"Let's use {torch.cuda.device_count()} GPUs!")
device = torch.device("cuda")
if torch.cuda.device_count() > 1:
stage1 = torch.nn.DataParallel(stage1)
diffusion = torch.nn.DataParallel(diffusion)
text_encoder = torch.nn.DataParallel(text_encoder)
stage1 = stage1.to(device)
diffusion = diffusion.to(device)
text_encoder = text_encoder.to(device)
optimizer = optim.AdamW(diffusion.parameters(), lr=config["ldm"]["base_lr"])
# Get Checkpoint
best_loss = float("inf")
start_epoch = 0
if resume:
print(f"Using checkpoint!")
checkpoint = torch.load(str(run_dir / "checkpoint.pth"))
diffusion.load_state_dict(checkpoint["diffusion"])
# Issue loading optimizer https://github.com/pytorch/pytorch/issues/2830
optimizer.load_state_dict(checkpoint["optimizer"])
start_epoch = checkpoint["epoch"]
best_loss = checkpoint["best_loss"]
else:
print(f"No checkpoint found.")
# Train model
print(f"Starting Training")
val_loss = train_ldm(
model=diffusion,
stage1=stage1,
scheduler=scheduler,
text_encoder=text_encoder,
start_epoch=start_epoch,
best_loss=best_loss,
train_loader=train_loader,
val_loader=val_loader,
optimizer=optimizer,
n_epochs=args.n_epochs,
eval_freq=args.eval_freq,
writer_train=writer_train,
writer_val=writer_val,
device=device,
run_dir=run_dir,
scale_factor=args.scale_factor,
)
log_mlflow(
model=diffusion,
config=config,
args=args,
experiment=args.experiment,
run_dir=run_dir,
val_loss=val_loss,
)
if __name__ == "__main__":
args = parse_args()
main(args)