-
Notifications
You must be signed in to change notification settings - Fork 0
/
run.py
211 lines (175 loc) · 9.23 KB
/
run.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
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
from pathlib import Path
import os
import losses
from model.ema import ExponentialMovingAverage
from model.ncsnpp import NCSNpp
import torch
import torch.optim as optim
from losses import get_sde_loss_fn
import numpy as np
import sde_lib
from model import utils
import logging
import datasets.datasets as data_loader
import time
import sampling
from torchvision.utils import make_grid, save_image
from semantic_synthesis.semantic_synthesis import get_semantic_synthesis_sampler
from semantic_synthesis.models.unet.unet import UNet
import datasets.cityscapes256.cityscapes256 as cityscapes256
import datasets.flickr.flickr as flickr
def train(config, workdir):
'''
Runs the training
:param config: The configuration
:param workdir: The working directory
'''
# Create sample directory
sample_dir = os.path.join(workdir, 'samples')
Path(sample_dir).mkdir(parents=True, exist_ok=True)
# Initialize models and optimizer
score_model = NCSNpp(config)
score_model = score_model.to(config.device)
score_model = torch.nn.DataParallel(score_model)
optimizer = optim.Adam(score_model.parameters(), lr=config.optim.lr, betas=(config.optim.beta1, 0.999),
eps=config.optim.eps, weight_decay=config.optim.weight_decay)
ema = ExponentialMovingAverage(score_model.parameters(), decay=config.model.ema_rate)
epoch = 1
logging.info('Model, EMA and optimizer initialized')
# Create checkpoint directories
checkpoint_dir = os.path.join(workdir, 'checkpoints')
Path(checkpoint_dir).mkdir(parents=True, exist_ok=True)
# Check for latest checkpoint
if os.path.isfile(os.path.join(checkpoint_dir, 'curr_cpt.pth')):
utils.restore_checkpoint(optimizer, score_model, ema, os.path.join(checkpoint_dir, 'curr_cpt.pth'))
logging.info('Checkpoint restored')
# Get data iterators
data_loader_train, data_loader_eval = data_loader.get_dataset(config)
logging.info('Dataset initialized')
# Get SDE
sde = sde_lib.get_SDE(config)
logging.info('SDE initialized')
# Get SDE loss function
loss_fn_train = get_sde_loss_fn(sde, True, reduce_mean=config.training.reduce_mean)
loss_fn_eval = get_sde_loss_fn(sde, False, reduce_mean=config.training.reduce_mean)
logging.info('Loss function loaded')
#Get step function
scaler = None if not config.optim.mixed_prec else torch.cuda.amp.GradScaler()
step_fn = losses.get_step_fn(config, score_model, optimizer, loss_fn_train, ema, scaler)
# Get sampling function
if config.training.snapshot_sampling:
sampling_shape = (config.sampling.batch_size, config.data.n_channels,
config.sampling.sampling_height, config.sampling.sampling_width)
sampling_fn = sampling.get_sampling_fn(config, sde, sampling_shape, config.model.sampling_eps)
logging.info(f'Starting training loop at epoch {epoch}')
step = 0
loss_per_log_period = 0
# Training cycle for one epoch
for i in range(epoch, config.training.epochs + 1):
start_time = time.time()
for img, _ in data_loader_train:
img = img.to(config.device)
# Training step
loss = step_fn(img, step)
step += 1
# Report loss and save to file
loss_per_log_period += loss
if step % config.training.log_freq == 0:
mean_loss = loss_per_log_period / config.training.log_freq
with open(os.path.join(workdir, 'training_loss.txt'), 'a+') as training_loss_file:
training_loss_file.write(str(step) + '\t' + str(mean_loss) + '\n')
logging.info(f'step: {step} (epoch: {epoch}), training_loss: {mean_loss}')
loss_per_log_period = 0
# Report the loss on an evaluation dataset and save to file
if step % config.training.eval_freq == 0 and data_loader_eval is not None:
total_loss = 0
with torch.no_grad():
ema.store(score_model.parameters())
ema.copy_to(score_model.parameters())
for eval_img, _ in data_loader_eval:
eval_img = eval_img.to(config.device)
eval_loss = loss_fn_eval(score_model, eval_img)
total_loss += eval_loss.item()
ema.restore(score_model.parameters())
total_loss = total_loss / len(data_loader_eval)
with open(os.path.join(workdir, 'eval_loss.txt'), 'a+') as eval_file:
eval_file.write(str(step) + '\t' + str(total_loss) + '\n')
logging.info(f'step: {step} (epoch: {epoch}), eval_loss: {total_loss}')
# Save the checkpoint
logging.info(f'Saving checkpoint of epoch {epoch}')
utils.save_checkpoint(optimizer, score_model, ema, epoch,
os.path.join(checkpoint_dir, 'curr_cpt.pth'))
if epoch % config.training.checkpoint_save_freq == 0:
utils.save_checkpoint(optimizer, score_model, ema, epoch,
os.path.join(checkpoint_dir, f'checkpoint_{epoch}.pth'))
# Generate and save samples
if config.training.snapshot_sampling and epoch % config.training.sampling_freq == 0:
ema.store(score_model.parameters())
ema.copy_to(score_model.parameters())
samples, n = sampling_fn(score_model)
ema.restore(score_model.parameters())
this_sample_dir = os.path.join(sample_dir, f'epoch_{epoch}')
Path(this_sample_dir).mkdir(parents=True, exist_ok=True)
nrow = int(np.sqrt(samples.shape[0]))
image_grid = make_grid(samples, nrow, padding=2)
save_image(image_grid, os.path.join(this_sample_dir, 'sample.png'))
logging.info(f'Samples generated in {this_sample_dir}')
time_for_epoch = time.time() - start_time
logging.info(f'Finished epoch {epoch}/{config.training.epochs} ({step // epoch} steps in this epoch) in {time_for_epoch} seconds')
epoch += 1
def image_synthesis(config, workdir, mode):
# Get checkpoint dir
checkpoint_dir = os.path.join(workdir, 'checkpoints')
# Create directory to sample_folder
sample_dir = os.path.join(workdir, 'sem_sample')
Path(sample_dir).mkdir(parents=True, exist_ok=True)
# Load score model from latest checkpoint
score_model = NCSNpp(config)
score_model = score_model.to(config.device)
score_model = torch.nn.DataParallel(score_model)
ema = ExponentialMovingAverage(score_model.parameters(), decay=config.model.ema_rate)
utils.restore_checkpoint(None, score_model, ema, os.path.join(checkpoint_dir, 'curr_cpt.pth'))
ema.copy_to(score_model.parameters())
logging.info('Score Model loaded')
# Load semantic segmentation model from checkpoint
sem_seg_model = UNet(config)
sem_seg_model = sem_seg_model.to(config.device)
sem_seg_model = torch.nn.DataParallel(sem_seg_model)
utils.restore_checkpoint(None, sem_seg_model, None, config.sampling.sem_seg_model_dir)
logging.info('Semantic Segmentation Model loaded')
# Get SDE
sde = sde_lib.get_SDE(config)
logging.info('SDE initialized')
if mode == 'cond':
# Get data iterators
data_loader_sample = data_loader.get_semantic_sample_data(config)
logging.info('Sample data loaded')
for i, (img, target, file_name) in enumerate(data_loader_sample):
img, target = img.to(config.device), target.to(config.device, dtype=torch.float32)
file_name = ''.join(file_name)
# Get sampling function
sampling_fn = get_semantic_synthesis_sampler(config, sde, score_model, sem_seg_model, target)
samples = sampling_fn()
nrow = int(np.sqrt(samples.shape[0]))
image_grid = make_grid(samples, nrow, padding=2, normalize=True)
save_image(image_grid, os.path.join(sample_dir, file_name + '_sample_0.000025_0.7_0.4_8labels_1000steps.png'))
save_image(img, os.path.join(sample_dir, file_name + '_original.png'))
# Save original mask as color image
if config.data.dataset == 'cityscapes256':
cityscapes256.save_colorful_images(target, sample_dir, file_name + '_mask.png')
if config.data.dataset == 'ade20k':
pass
#ade20k.save_colorful_images(target, sample_dir, file_name + '_mask.png')
if config.data.dataset == 'flickr':
flickr.save_colorful_images(target, sample_dir, file_name + '_mask.png')
logging.info(f'Generated sample {i + 1} of {len(data_loader_sample)}')
elif mode == 'uncond':
sampling_shape = (config.sampling.batch_size, config.data.n_channels,
config.sampling.sampling_height, config.sampling.sampling_width)
sampling_fn = sampling.get_sampling_fn(config, sde, sampling_shape, config.model.sampling_eps)
for i in range(0, 20):
samples, n = sampling_fn(score_model)
nrow = int(np.sqrt(samples.shape[0]))
image_grid = make_grid(samples, nrow, padding=2, normalize=True)
save_image(image_grid, os.path.join(sample_dir, f'{i}_uncond_sample.png'))
logging.info(f'Generated sample {i + 1} of {20}')