-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_nsgan.py
288 lines (240 loc) · 13.1 KB
/
train_nsgan.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
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
from determined.pytorch import PyTorchTrial, PyTorchTrialContext, DataLoader
from determined.tensorboard.metric_writers.pytorch import TorchWriter
from models.nsgan import Generator, Discriminator
from torch.optim import Adam
import torch
import torch.nn as nn
from torchvision.utils import make_grid
from datasets import Datasets
from typing import Dict, Sequence, Union
from metrics.fid_infinity import FIDInfinity
from metrics.is_infinity import ISInfinity
from torchmetrics import FID, IS, KID
from utility.utils import parse_none_true_false, adjust_dimensions, normalize_to_0_255, normalize_to_0_1, save_images_to_disk
from torchmetrics.utilities.data import dim_zero_cat
import yaml
import math
TorchData = Union[Dict[str, torch.Tensor], Sequence[torch.Tensor], torch.Tensor]
class TrainNSGAN(PyTorchTrial):
"""
Entry point for training a Non Saturating Generative Adversarial Network.
Initializes generator and discriminator network, loss function, metrics and optimizers.
Parses configuration from associated config file and global config file.
Args:
context:
A PyTorchTrialContext object
"""
def __init__(self, context: PyTorchTrialContext) -> None:
self.context = context
self.data_config = self.context.get_data_config()
self.img_dim = tuple(self.data_config['dim'])
self.dataset_name = self.data_config['dataset']
self.global_parameters = self.data_config['global_parameters']
# Load Hyperparameters from config file
self.batch_size = self.context.get_per_slot_batch_size()
self.learning_rate = self.context.get_hparam('lr')
self.beta1 = self.context.get_hparam('b1')
self.beta2 = self.context.get_hparam('b2')
self.latent_dim = self.context.get_hparam('latent_dim')
self.disc_iterations = self.context.get_hparam('disc_iterations')
self.evaluate_while_trainig = self.context.get_hparam('evaluate_while_trainig')
with open(self.data_config['global_parameters'], 'r') as file:
try:
data = yaml.safe_load(file)
if self.evaluate_while_trainig:
self.training_metrics = data['training_metrics']
self.evaluation_metrics = data['evaluation_metrics']
self.save_images_config = data['save_images']
except yaml.YAMLError:
print(f"Couldn't load metrics from {self.data_config['global_parameters']}")
# Map strings to class instances
if self.evaluate_while_trainig:
self.training_metrics = list(map(lambda data: self.context.to_device(globals()[data[0]](**parse_none_true_false(data[1]))), self.training_metrics.items()))
self.save_real_images = self.save_images_config['real_images']['save']
self.save_fake_images = self.save_images_config['fake_images']['save']
if self.save_real_images:
self.number_of_real_images_to_save = self.save_images_config['real_images']['number_of_images']
self.real_images_to_save_path = self.save_images_config['real_images']['path']
self.real_images_to_save_name = self.save_images_config['real_images']['file_name']
if self.save_fake_images:
self.number_of_fake_images_to_save = self.save_images_config['fake_images']['number_of_images']
self.fake_images_to_save_path = self.save_images_config['fake_images']['path']
self.fake_images_to_save_name = self.save_images_config['fake_images']['file_name']
self.dataset_directory = f"/tmp/data-rank{self.context.distributed.get_rank()}"
self.datasets = Datasets(self.dataset_directory, self.batch_size)
self.tb_writer = TorchWriter().writer
# Initialize Networks
self.generator = self.context.wrap_model(Generator(self.latent_dim, self.img_dim, self.batch_size))
self.discriminator = self.context.wrap_model(Discriminator(self.img_dim, self.batch_size))
# Initialize Optimizers
self.optimizer_gen = self.context.wrap_optimizer(Adam(self.generator.parameters(), lr=self.learning_rate, betas=[self.beta1, self.beta2]))
self.optimizer_disc = self.context.wrap_optimizer(Adam(self.discriminator.parameters(), lr=self.learning_rate, betas=[self.beta1, self.beta2]))
# Initialize Loss Function
self.loss_func = nn.BCELoss()
if self.evaluate_while_trainig:
if self.evaluate_while_trainig:
self.training_scores = dict()
for m in self.training_metrics:
self.training_scores[type(m).__name__] = float("NaN")
def train_batch(self, batch: TorchData, epoch_idx: int, batch_idx: int) -> Dict[str, torch.Tensor]:
'''Calculate training losses and metrics on a single batch. Periodically writes data to TensorBoard.
Args:
batch:
A single batch containing training data and labels
epoch_idx:
Index of training epoch
batch_idx:
Index of batch
Returns:
Dictionary containing calculated losses and metrics for this particular batch
'''
real_imgs, _ = batch
self.generator.requires_grad_(True)
self.discriminator.requires_grad_(False)
# Create random noise
random_noise = self.context.to_device(torch.randn(self.batch_size, self.latent_dim))
# Generate Images
fake_imgs = self.generator(random_noise)
# Train Generator only every self.disc_iterations iterations
if batch_idx % self.disc_iterations == 0:
ones = self.context.to_device(torch.ones(self.batch_size, 1))
self.generator_loss = self.loss_func(self.discriminator(fake_imgs), ones)
self.context.backward(self.generator_loss)
self.context.step_optimizer(self.optimizer_gen)
#Train Discriminator
self.generator.requires_grad_(False)
self.discriminator.requires_grad_(True)
ones = self.context.to_device(torch.ones(self.batch_size, 1))
loss_real_imgs = self.loss_func(self.discriminator(real_imgs), ones)
zeros = self.context.to_device(torch.zeros(self.batch_size, 1))
loss_fake_imgs = self.loss_func(self.discriminator(fake_imgs.detach()), zeros)
discriminator_loss = (loss_real_imgs + loss_fake_imgs) / 2
self.context.backward(discriminator_loss)
self.context.step_optimizer(self.optimizer_disc)
# Write to Tensorboard
if batch_idx % 100 == 0:
tb_imgs = fake_imgs.view(self.batch_size, *self.img_dim)[0:16]
tb_imgs = adjust_dimensions(tb_imgs)
tb_imgs = normalize_to_0_255(tb_imgs).to(dtype=torch.uint8)
img_grid = make_grid(tb_imgs)
self.tb_writer.add_image(f'raw: epoch index: {epoch_idx}, batch index: {batch_idx}', img_grid)
if self.evaluate_while_trainig:
with torch.no_grad():
real_imgs, fake_imgs = adjust_dimensions(real_imgs), adjust_dimensions(fake_imgs)
real_imgs_0_255, fake_imgs_0_255 = normalize_to_0_255(real_imgs).to(dtype=torch.uint8), normalize_to_0_255(fake_imgs).to(dtype=torch.uint8)
real_imgs_0_1, fake_imgs_0_1 = normalize_to_0_1(real_imgs), normalize_to_0_1(fake_imgs)
for m in self.training_metrics:
if isinstance(m, (FID, KID)):
# 0-255, uint8
m.update(real_imgs_0_255, real=True)
m.update(fake_imgs_0_255, real=False)
elif isinstance(m, FIDInfinity):
# 0-1, float32
m.update(real_imgs_0_1, real=True)
m.update(fake_imgs_0_1, real=False)
elif isinstance(m, ISInfinity):
# 0-1, float32
m.update(fake_imgs_0_1)
else:
# 0-255, uint8
m.update(fake_imgs_0_255)
for m in self.training_metrics:
if isinstance(m, KID):
if len(dim_zero_cat(m.real_features)) >= 5000:
score, _ = m.compute()
else:
score = float("NaN")
elif isinstance(m, (ISInfinity, FIDInfinity)):
if m.get_number_of_features() >= 5000:
score = m.compute()
else:
score = float("NaN")
elif isinstance(m, IS):
score, _ = m.compute()
else:
score = m.compute()
self.training_scores[type(m).__name__] = score
losses = {
"generator_loss": self.generator_loss,
"discriminator_loss": discriminator_loss,
}
if self.evaluate_while_trainig:
return {**self.training_scores, **losses}
else:
return losses
def evaluate_full_dataset(self, data_loader: DataLoader) -> Dict[str, torch.Tensor]:
'''Calculate metrics on full evaluation dataset. Save images to disk if activated in settings.
Args:
data_loader:
DataLoader containing evaluation dataset
Returns:
Dictionary containing calculated metrics for the evaluation dataset
'''
real_data_to_disk = []
evaluation_metrics = list(map(lambda data: self.context.to_device(globals()[data[0]](**parse_none_true_false(data[1]))), self.evaluation_metrics.items()))
scores = dict()
for imgs, _ in data_loader:
real_imgs = self.context.to_device(imgs)
# Create random noise
random_noise = self.context.to_device(torch.randn(self.batch_size, self.latent_dim))
# Generate Images
fake_imgs = self.generator(random_noise)
real_imgs, fake_imgs = adjust_dimensions(real_imgs), adjust_dimensions(fake_imgs)
real_imgs_0_255, fake_imgs_0_255 = normalize_to_0_255(real_imgs).to(dtype=torch.uint8), normalize_to_0_255(fake_imgs).to(dtype=torch.uint8)
real_imgs_0_1, fake_imgs_0_1 = normalize_to_0_1(real_imgs), normalize_to_0_1(fake_imgs)
if self.save_real_images:
if len(real_data_to_disk) == 0 or len(dim_zero_cat(real_data_to_disk)) < self.number_of_real_images_to_save:
real_data_to_disk.append(real_imgs_0_255)
for m in evaluation_metrics:
if isinstance(m, (FID, KID)):
# 0-255, uint8
m.update(real_imgs_0_255, real=True)
m.update(fake_imgs_0_255, real=False)
elif isinstance(m, FIDInfinity):
# 0-1, default
m.update(real_imgs_0_1, real=True)
m.update(fake_imgs_0_1, real=False)
elif isinstance(m, ISInfinity):
# 0-1, default
m.update(fake_imgs_0_1)
else:
# 0-255, uint8
m.update(fake_imgs_0_255)
for m in evaluation_metrics:
if isinstance(m, KID):
if len(dim_zero_cat(m.real_features)) >= 5000:
score, _ = m.compute()
else:
score = float("NaN")
elif isinstance(m, (ISInfinity, FIDInfinity)):
if m.get_number_of_features() >= 5000:
score = m.compute()
else:
score = float("NaN")
elif isinstance(m, (IS, KID)):
score, _ = m.compute()
else:
score = m.compute()
if not math.isnan(score):
scores[type(m).__name__] = score
if self.save_fake_images:
disk_imgs = self.generator(self.context.to_device(torch.randn(self.number_of_fake_images_to_save, self.latent_dim)), shape=[self.number_of_fake_images_to_save, *self.img_dim])
disk_imgs = adjust_dimensions(disk_imgs)
disk_imgs = normalize_to_0_255(disk_imgs).to(dtype=torch.uint8)
save_images_to_disk(disk_imgs, path=self.fake_images_to_save_path, file_name=self.fake_images_to_save_name)
if self.save_real_images:
real_data_to_disk = dim_zero_cat(real_data_to_disk)[:self.number_of_real_images_to_save]
save_images_to_disk(real_data_to_disk, path=self.real_images_to_save_path, file_name=self.real_images_to_save_name)
return scores
def build_training_data_loader(self) -> DataLoader:
"""Build DataLoader for training dataset
Returns:
DataLoader
"""
return self.datasets.get_data_loader(self.dataset_name, train=True)
def build_validation_data_loader(self) -> DataLoader:
"""Build DataLoader for evaluation dataset
Returns:
DataLoader
"""
return self.datasets.get_data_loader(self.dataset_name, train=False)