-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrainer.py
179 lines (148 loc) · 7.99 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
from typing import List
import torch
import torch.nn as nn
import torch.nn.functional as F
import os
import copy
import torch
from torch import nn, optim
from torch.utils.tensorboard import SummaryWriter
from tqdm.notebook import tqdm
from vqgan.interface import pretrained_vqgan
from data import load_data
from diffusion import Diffusion
from unet import UNet
import pytorch_lightning as pl
from torchvision.utils import make_grid
from argparse import Namespace
from text_encode import SentenceEmbedder
class EMA:
def __init__(self, beta):
self.beta = beta
self.step = 0
def step_ema(self, model, ema_model, start_step = 50):
self.step += 1
if self.step < start_step:
ema_model.load_state_dict(model.state_dict())
return
for current_param, ema_param in zip(model.parameters(), ema_model.parameters()):
current_weight, ema_weight = current_param.data, ema_param.data
ema_param.data = self.beta * ema_weight + (1 - self.beta) * current_weight
class DiffusionTrainer(pl.LightningModule):
def __init__(self, args):
super(DiffusionTrainer, self).__init__()
self.args = args
print(self.device)
self.diffusion = Diffusion(cosine = args.cosine_scheduler, image_size=args.image_size, device=args.device)
self.vqgan = pretrained_vqgan().eval()
for param in self.vqgan.parameters():
param.requires_grad = False
self.encode = self.vqgan.encode
self.decode = self.vqgan.decode
self.ema = EMA(0.995)
self.model = UNet()
self.ema_model = copy.deepcopy(self.model).eval().requires_grad_(False)
self.se = SentenceEmbedder()
def forward(self, x_t, t):
return self.model(x_t, t)
def configure_optimizers(self):
return optim.AdamW(self.model.parameters(), lr=self.args.lr)
def log_images(self, noised_image, reconstructed_image, orig_image, predicted_noise, step, stage, image_type):
noised_image, reconstructed_image, orig_image, predicted_noise = noised_image.mul(0.5).add(0.5), reconstructed_image.mul(0.5).add(0.5), orig_image.mul(0.5).add(0.5), predicted_noise.mul(0.5).add(0.5)
grid = make_grid(torch.cat([noised_image, reconstructed_image, orig_image, predicted_noise]), nrow=2)
self.logger.experiment.add_image(f"{stage}/{image_type}_Images", grid, global_step=step)
def on_fit_start(self, *args, **kwargs):
super().on_fit_start(*args, **kwargs)
self.diffusion.device = self.device
self.diffusion.noise_schedule()
def training_step(self, batch, batch_idx):
images, labels = batch
captions = torch.from_numpy(self.se.encode_sentences(labels['TEXT'])) if self.args.conditional else None
t = torch.randint(low=1, high=self.diffusion.num_steps, size=(images.shape[0],)).to(self.device)
encoded_images = self.encode(images)
x_t, noise = self.diffusion.forward(encoded_images, t)
predicted_noise = self.model(x_t, t, captions)
loss = nn.MSELoss()(noise, predicted_noise)
# Log images
if batch_idx % self.args.log_image_interval == 0:
predicted_images = x_t - predicted_noise
pixel_equiv_noised_image = self.decode(x_t)
pixel_reconstructed_images = self.decode(predicted_images) # Get the reconstructed image
pixel_predicted_noise = images - pixel_reconstructed_images
self.log_images(pixel_equiv_noised_image, pixel_reconstructed_images, images, pixel_predicted_noise, self.global_step, "Train", "Pixel-Space")
self.log_images(x_t, predicted_images, encoded_images, noise, self.global_step, "Train", "Latent-Space")
self.log("Train/MSE", loss)
self.log("mse", loss, prog_bar=True, logger=False) # Add this line
return {"loss": loss}
def optimizer_step(self, *args, **kwargs):
super().optimizer_step(*args, **kwargs)
self.ema.step_ema(self.model, self.ema_model)
def on_train_epoch_end(self):
self.diffusion.device = self.device
if self.device == torch.device('cuda:0'):
with torch.no_grad():
sampled_images = self.diffusion.sample_n_images(self.vqgan, self.model, 2)
ema_sampled_images = self.diffusion.sample_n_images(self.vqgan, self.ema_model, 2)
self.logger.experiment.add_images("EOE samples - pixel", sampled_images, self.trainer.current_epoch, dataformats = 'NCHW')
self.logger.experiment.add_images("EOE EMA samples - pixel", ema_sampled_images, self.trainer.current_epoch, dataformats = 'NCHW')
def validation_step(self, batch, batch_idx):
images, labels = batch
captions = torch.from_numpy(self.se.encode_sentences(labels['TEXT'])) if self.args.conditional else None
t = torch.randint(low=1, high=self.diffusion.num_steps, size=(images.shape[0],)).to(self.device)
encoded_images = self.encode(images)
x_t, noise = self.diffusion.forward(encoded_images, t)
predicted_noise = self.model(x_t, t. captions)
loss = nn.MSELoss()(noise, predicted_noise)
# Log images
if batch_idx % self.args.log_image_interval == 0:
predicted_images = x_t - predicted_noise
pixel_equiv_noised_image = self.decode(x_t)
pixel_reconstructed_images = self.decode(predicted_images)
pixel_predicted_noise = images - pixel_reconstructed_images
self.log_images(pixel_equiv_noised_image, pixel_reconstructed_images, images, pixel_predicted_noise, self.global_step, "Val", "Pixel-Space")
self.log_images(x_t, predicted_images, encoded_images, noise, self.global_step, "Val", "Latent-Space")
self.log("Val/MSE", loss)
self.log("mse", loss, prog_bar=True, logger=False) # Add this line
return {"val_loss": loss}
def train_dataloader(self):
return load_data(self.args.dataset_path, self.args.batch_size, self.args.orig_resolution, self.args.num_workers)
def val_dataloader(self):
return load_data(self.args.dataset_path)
default_args = Namespace(
run_name='diff',
orig_resolution = 256,
epochs=30,
batch_size=4,
dataset_path='wikiart_images',
device='cuda',
lr=3e-4,
save_interval=100,
image_size=[32, 32],
log_image_interval=10,
num_workers = 4
)
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--run_name', default='diff', type=str)
parser.add_argument('--orig_resolution', default=256, type=int)
parser.add_argument('--epochs', default=10, type=int)
parser.add_argument('--batch_size', default=8, type=int)
parser.add_argument('--dataset_path', default='wikiart_images', type=str)
parser.add_argument('--val_dataset_path', default='wikiart_images', type=str)
parser.add_argument('--device', default='cuda', type=str)
parser.add_argument('--lr', default=3e-4, type=float)
parser.add_argument('--val_batch_per_epoch', default=0, type=int)
parser.add_argument('--save_interval', default=100, type=int)
parser.add_argument('--image_size', default=[32, 32], type=int, nargs=2)
parser.add_argument('--log_image_interval', default=50, type=int)
parser.add_argument('--num_workers', default=0, type=int)
parser.add_argument('--checkpoint_path', default=None, type=str)
parser.add_argument('--train_batches_per_epoch', default = 1000, type = int)
parser.add_argument('--cosine_scheduler', default = False, type = bool)
parser.add_argument('--conditional', default = False, type=bool)
args = parser.parse_args()
# Example of how to train the DiffusionTrainer using PyTorch Lightning
trainer = pl.Trainer(limit_train_batches = args.train_batches_per_epoch, max_epochs=args.epochs, accelerator = 'gpu', limit_val_batches=args.val_batch_per_epoch, log_every_n_steps=1)
diffusion_trainer = DiffusionTrainer(args)
trainer.fit(diffusion_trainer, ckpt_path = args.checkpoint_path)