-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrainer.py
190 lines (151 loc) · 6.56 KB
/
trainer.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
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
from dataset import FlowDataset
from model import FlowModel
from diffusers.models import AutoencoderKL
from transformers import CLIPTokenizer, CLIPTextModel
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import einops
import numpy as np
from PIL import Image
import yaml
import argparse
import os
class FlowTrainer():
def __init__(self, config_path, device):
self.config_path = config_path
self.device = device
with open(self.config_path, 'r') as file:
config = yaml.load(file, Loader=yaml.FullLoader)
self.dataset_config = config["dataset"]
self.model_config = config["model"]
self.train_config = config["train"]
self.eval_config = config["eval"]
self.dataset = FlowDataset(self.dataset_config["data_folder"])
self.dataloader = DataLoader(self.dataset, batch_size=self.dataset_config["batch_size"], num_workers=4)
self.vae = AutoencoderKL.from_pretrained(
"CompVis/stable-diffusion-v1-4",
subfolder="vae",
revision="fp16",
torch_dtype="auto"
)
self.vae.to(device)
self.vae.eval()
for param in self.vae.parameters():
param.requires_grad = False
self.clip_tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
self.clip_text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
self.clip_text_encoder.to(device)
for param in self.clip_text_encoder.parameters():
param.requires_grad = False
self.model = FlowModel(
n_blocks=self.model_config["n_blocks"],
c_latent=self.model_config["c_latent"],
d_embd=self.model_config["d_embd"],
n_head=self.model_config["n_head"],
d_time_embd=self.model_config["d_time_embd"],
d_cond_embd=self.clip_text_encoder.config.hidden_size,
patch_size=self.model_config["patch_size"],
)
self.initialize_weights()
self.model.to(device)
self.optimizer = optim.AdamW(
self.model.parameters(),
lr=self.train_config["learning_rate"],
weight_decay=self.train_config["weight_decay"]
)
self.loss = nn.MSELoss()
def initialize_weights(self):
for name, module in self.model.named_modules():
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
def train_one_step(self, image, desc):
image = image.to(self.device)
b, c, h, w = image.shape
with torch.no_grad():
image_enc = self.vae.encode(image)
latent = image_enc.latent_dist.sample()
l_b, l_c, l_h, l_w = latent.shape
with torch.no_grad():
inputs = self.clip_tokenizer(
desc,
return_tensors='pt',
padding=True,
truncation=True
)
inputs = {k: v.to(self.device) for k, v in inputs.items()}
cond = self.clip_text_encoder(**inputs).last_hidden_state
noise = torch.randn(l_b, l_c, l_h, l_w).to(image.device)
time = torch.rand(b).to(image.device) ** 2
time = torch.clamp(time, min=1e-4)
time_broadcast = einops.rearrange(time, "b -> b 1 1 1")
# x = noise + t*(latent - noise) -> x + (1-t)*(latent-noise) = latent
x = time_broadcast * latent + (1 - time_broadcast)*noise
vf_pred = self.model(x, time, cond)
loss = self.loss(latent - noise, vf_pred)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
return loss
def train(self):
for i in range(self.train_config["n_epochs"]):
for k, batch in enumerate(self.dataloader):
loss = self.train_one_step(batch[0], batch[1])
print("Iteration:", k, "Loss:", loss.detach().cpu().numpy())
if k % self.eval_config["iters"] == 0:
self.eval(k ,batch[0][0:2], batch[1][0:2])
if k % self.train_config["save_ckpt_interval"] == 0:
self.save_checkpoint(k)
@torch.no_grad()
def eval(self, num_iters, image, desc):
image = image.to(self.device)
b, c, h, w = image.shape
with torch.no_grad():
image_enc = self.vae.encode(image)
latent = image_enc.latent_dist.sample()
l_b, l_c, l_h, l_w = latent.shape
with torch.no_grad():
inputs = self.clip_tokenizer(
desc,
return_tensors='pt',
padding=True,
truncation=True
)
inputs = {k: v.to(self.device) for k, v in inputs.items()}
cond = self.clip_text_encoder(**inputs).last_hidden_state
latent = torch.randn(l_b, l_c, l_h, l_w, device=image.device)
time_step = torch.tensor([1 / self.eval_config["steps"]], device=image.device)
for i in range(self.eval_config["steps"]):
time = torch.tensor([i / self.eval_config["steps"]], device=image.device)
vf_pred = self.model(latent, time, cond)
latent += vf_pred * time_step
dec_outputs = self.vae.decode(latent)
dec_images = dec_outputs.sample
dec_images = einops.rearrange(dec_images, "b c h w -> b h w c")
dec_images = dec_images.detach().cpu().numpy()
dec_images = np.clip(dec_images, 0, 1)
dec_images = np.array(dec_images*255, dtype=np.uint8)
eval_folder = self.eval_config["eval_folder"]
os.makedirs(eval_folder, exist_ok=True)
for i in range(len(dec_images)):
dec_image_pil = Image.fromarray(dec_images[i])
iter_folder = f"{eval_folder}/iteration_{num_iters}"
os.makedirs(iter_folder, exist_ok=True)
dec_image_pil.save(f"{iter_folder}/img_{i}.jpg")
def save_checkpoint(self, iterations):
checkpoint = {
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
}
ckpt_folder = self.train_config["ckpt_folder"]
os.makedirs(ckpt_folder, exist_ok=True)
filepath = f"{ckpt_folder}/iteration_{iterations}.pth"
torch.save(checkpoint, filepath)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--config_path", type=str)
args = parser.parse_args()
device = "cuda"
trainer = FlowTrainer(args.config_path, device)
trainer.train()