-
Notifications
You must be signed in to change notification settings - Fork 0
/
VAE_binary.py
355 lines (308 loc) · 16.9 KB
/
VAE_binary.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
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
import numpy as np
from models import VAE, Elbo_BCE
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from util import load_data
from torchvision.utils import save_image
import time
import os
import csv
import torchsummary
import io
from contextlib import redirect_stdout
from math import ceil
#K is the number of importance samples
#This functions loads the data from the loader, samples k z_i from the q(z|x_i) for each x_i,
#and calls evaluate_batch_likelihood function to get the likelihood using importance sampling
def estimate_data_likelihood(model, loader, K=200):
data_log_likelihood = []
print("Evaluating Likelihood .", end="")
with torch.no_grad():
for batch_id, (data, _) in enumerate(loader):
data = data.to(device)
mean, logvar = model.encoder(data)
mean = mean.unsqueeze(1).expand(-1, K, -1)
logvar = logvar.unsqueeze(1).expand(-1, K, -1)
z_samples = model.reparametrize(mean, logvar)
batch_log_likelihood = evaluate_batch_likelihood(model, data.reshape((data.shape[0], -1)), z_samples)
data_log_likelihood += batch_log_likelihood.tolist()
print(".", end="")
print("\n", end="")
return data_log_likelihood
#Evaluate batch log-likelihood using importance sampling
#x is of size (batch_size M, features_size D)
#z_samples is of size ((batch_size M, importance samples K, latent size L)
def evaluate_batch_likelihood(model, x, z_samples, image_size=(1,28,28)):
with torch.no_grad():
z_samples = z_samples.to(device)
x = x.to(device)
M = x.shape[0]; D = x.shape[1];
K = z_samples.shape[1]; L = z_samples.shape[2];
mean, logvar = model.encoder(x.reshape((M, image_size[0], image_size[1], image_size[2])))
std = torch.exp(logvar / 2)
x_tilde_logits = model.decoder(z_samples.reshape((M*K, L)))
x_tilde_logits = x_tilde_logits.reshape((M, K, D))
x = x.unsqueeze(1).expand(-1, K, -1)
mean = mean.unsqueeze(1).expand(-1, K, -1)
std = std.unsqueeze(1).expand(-1, K, -1)
log_encoder_zi = -0.5 * torch.sum(((z_samples - mean)/std)**2, -1) - 0.5 * torch.sum(torch.log(2*np.pi*std**2), -1)
log_prior_zi = -0.5 * torch.sum(z_samples ** 2, -1) - 0.5 * L * np.log(2 * np.pi)
log_decoder_zi = -torch.sum(F.binary_cross_entropy_with_logits(input=x_tilde_logits, target=x, reduction="none")
, dim=-1)
log_term = log_decoder_zi + log_prior_zi - log_encoder_zi
#LogSumExp trick
max_log_term, _ = log_term.max(dim=-1, keepdim=True)
batch_log_likelihood = -np.log(K) + max_log_term.squeeze(-1) + (log_term - max_log_term).exp().sum(-1).log()
return batch_log_likelihood
#Train for one epoch
def epoch_train(model, optimizer, loader, loss_fn, epoch, log_interval=None):
model.train()
elbo = []
for batch_id, (x, _) in enumerate(loader):
x = x.to(device)
optimizer.zero_grad()
x_tilde_logits, mean, logvar = model(x)
loss = loss_fn(x, x_tilde_logits, mean, logvar)
loss.backward()
optimizer.step()
elbo.append(-loss.item())
if log_interval is not None:
if batch_id%log_interval == 0:
print (f"----> Epoch {epoch},\t Batch {batch_id},\t Train ELBO: {np.mean(np.array(elbo)):.2f}")
return np.mean(np.array(elbo))
#Evaluate the model on a specific dataloader
def epoch_eval(model, loader, loss_fn):
model.eval()
elbo = []
for batch_id, (x, _) in enumerate(loader):
x = x.to(device)
x_tilde_logits, mean, logvar = model(x)
loss = loss_fn(x, x_tilde_logits, mean, logvar)
elbo.append(-loss.item())
return np.mean(np.array(elbo))
#Train the model for a number of epochs
def train(model, optimizer, train_loader, val_loader, loss_fn, epochs, save_dir = os.curdir, save_interval=None, log_interval=None,
model_outputs_logits=True, train_samples=None, val_samples=None, random_z=None):
train_elbos = []
val_elbos = []
epoch_time = []
max_val_elbo = -1000000
for epoch in range(epochs):
stime = time.time()
train_elbo = epoch_train(model, optimizer, train_loader, loss_fn, epoch, log_interval)
val_elbo = epoch_eval(model, val_loader, loss_fn)
train_elbos.append(train_elbo)
val_elbos.append(val_elbo)
epoch_time_ = time.time() - stime
epoch_time.append(epoch_time_)
if val_elbo > max_val_elbo:
max_val_elbo = val_elbo
if save_interval is not None:
save_model(model, optimizer, train_elbos, val_elbos, epoch_time, epoch, save_dir, True)
generate_samples(model, save_dir=save_dir, epoch=epoch, train_samples=train_samples,
val_samples=val_samples, random_z=random_z, model_outputs_logits=model_outputs_logits, best=True)
if save_interval is not None:
if epoch % save_interval == 0:
save_model(model, optimizer, train_elbos, val_elbos, epoch_time, epoch, save_dir, False)
generate_samples(model, save_dir=save_dir, epoch=epoch, train_samples=train_samples,
val_samples=val_samples, random_z=random_z, model_outputs_logits=model_outputs_logits)
print(f"-> Epoch {epoch},\t Train ELBO: {train_elbo:.2f},\t Validation ELBO: {val_elbo:.2f},\t "
f"Max Validation ELBO: {max_val_elbo:.2f},\t Epoch Time: {epoch_time_:.2f} seconds")
if save_interval is not None:
save_model(model, optimizer, train_elbos, val_elbos, epoch_time, epoch, save_dir, False)
generate_samples(model, save_dir=save_dir, epoch=epoch, train_samples=train_samples,
val_samples=val_samples, random_z=random_z, model_outputs_logits=model_outputs_logits)
return train_elbos, val_elbos, epoch_time[-1]
#Save the model and the training statistics
def save_model(model, optimizer, train_elbos, val_elbos, epoch_time, epoch, save_dir, best_model=False):
if best_model:
path = os.path.join(save_dir, f'model_best.pt')
else:
path = os.path.join(save_dir, f'model_epoch_{epoch}.pt')
torch.save({
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'train_elbos': train_elbos,
'val_elbos': val_elbos,
}, path)
epochs = [j for j in range(epoch+1)]
stats = {'Epoch': epochs, 'Train Elbo': train_elbos, "Validation Elbo": val_elbos, "Epoch Time": epoch_time}
stats_path = os.path.join(save_dir, 'stats.csv')
with open(stats_path, 'w') as csvfile:
fieldnames = stats.keys()
writer = csv.writer(csvfile)
writer.writerow(fieldnames)
writer.writerows([[stats[key][j] for key in fieldnames] for j in epochs])
#Print the model architecture and hyperparameters
def print_model_summary(model, optimizer, save_dir=None, input_size=(1,28,28)):
f = io.StringIO()
with redirect_stdout(f):
print(optimizer)
torchsummary.summary(model, input_size)
architecture_summary = f.getvalue()
print(architecture_summary)
if save_dir is not None:
architecture_path = os.path.join(save_dir, 'architecture.txt')
with open(architecture_path, 'w') as file:
file.write(architecture_summary)
#Generate original samples vs reconstructed samples for the training and validation sets, plus some random samples
def generate_samples(model, save_dir, epoch, train_samples, val_samples, random_z, model_outputs_logits=True, best=False):
with torch.no_grad():
train_samples = train_samples.to(device)
val_samples = val_samples.to(device)
random_z = random_z.to(device)
generated_train_samples, mean,_ = model(train_samples)
generated_val_samples, _, _ = model(val_samples)
generated_random_samples = model.decoder(random_z)
if model_outputs_logits:
generated_train_samples = (torch.sigmoid(generated_train_samples)).round().cpu()
generated_val_samples = (torch.sigmoid(generated_val_samples)).round().cpu()
generated_random_samples = (torch.sigmoid(generated_random_samples)).round().cpu()
else: #assuming input images are in the range of -1 and 1
generated_train_samples = ((generated_train_samples + 1) / 2).cpu()
generated_val_samples = ((generated_val_samples + 1) / 2).cpu()
generated_random_samples = ((generated_random_samples + 1) / 2).cpu()
if epoch == 0:
if model_outputs_logits:
save_image(train_samples, os.path.join(save_dir, "original_train_samples.png"))
save_image(val_samples, os.path.join(save_dir, "original_val_samples.png"))
else: # assuming input images are in the range of -1 and 1
save_image((train_samples + 1) / 2, os.path.join(save_dir, "original_train_samples.png"))
save_image((val_samples + 1) / 2, os.path.join(save_dir, "original_val_samples.png"))
if best:
generated_train_path = os.path.join(save_dir, f"generated_train_samples_best.png")
generated_val_path = os.path.join(save_dir, f"generated_val_samples_best.png")
generated_random_path = os.path.join(save_dir, f"generated_random_samples_best.png")
else:
generated_train_path = os.path.join(save_dir, f"generated_train_samples_{epoch}.png")
generated_val_path = os.path.join(save_dir, f"generated_val_samples_{epoch}.png")
generated_random_path = os.path.join(save_dir, f"generated_random_samples_{epoch}.png")
save_image(generated_train_samples, generated_train_path)
save_image(generated_val_samples, generated_val_path)
save_image(generated_random_samples, generated_random_path)
#Generate a num_samples random samples
def generate_random_samples(model, save_dir, random_z_samples=None, epoch=-1, num_samples=200, latent_size=100,
model_outputs_logits=True, samples_per_image=64, generated_random_file="generated_random_samples"):
with torch.no_grad():
num_images = ceil(num_samples / samples_per_image)
for k in range(1, num_images +1):
if k == num_images and num_samples%samples_per_image != 0:
samples_per_image = num_samples%samples_per_image
if random_z_samples is None:
random_z = torch.randn((samples_per_image, latent_size))
else:
try:
random_z = random_z_samples[(k-1)*samples_per_image:k*samples_per_image]
except:
if len(random_z_samples) < k*samples_per_image:
print("random_z_samples are too small for the number of images requested")
random_z = random_z.to(device)
generated_random_samples = model.decoder(random_z)
if model_outputs_logits:
generated_random_samples = (torch.sigmoid(generated_random_samples)).round().cpu()
else: #assuming input images are in the range of -1 and 1
generated_random_samples = ((generated_random_samples + 1) / 2).cpu()
generated_random_file_ = generated_random_file + f"_epoch_{epoch}_numsamples_{num_samples}_{k:03d}.png"
image_path = os.path.join(save_dir, generated_random_file_)
save_image(generated_random_samples, image_path)
#Generate an interpolated image between two images generated given two latent variables
#random_z_1 and random_z_2 with factor alpha x = alpha*x_0 + (a-alpha)*x_1
def generate_interpolated_samples(model, save_dir, alphas, interpolate_images=False, random_z_0=None, random_z_1=None, epoch=-1, num_samples=64, latent_size=100,
model_outputs_logits=False, generated_random_file="generated_random_samples_interpolated"):
save_dir = os.path.join(save_dir, "interpolation_samples")
if not os.path.isdir(save_dir):
os.mkdir(save_dir)
with torch.no_grad():
if (random_z_0 is None) or (random_z_1 is None):
random_z_0 = torch.randn((num_samples, latent_size))
random_z_1 = torch.randn((num_samples, latent_size))
random_z_0 = random_z_0.to(device)
random_z_1 = random_z_1.to(device)
if interpolate_images:
generated_images_0 = model.decoder(random_z_0)
generated_images_1 = model.decoder(random_z_1)
for alpha in alphas:
generated_images = alpha*generated_images_0 + (1.-alpha) * generated_images_1
if model_outputs_logits:
generated_images = (torch.sigmoid(generated_images)).round().cpu()
else: # assuming input images are in the range of -1 and 1
generated_images = ((generated_images + 1) / 2).cpu()
generated_random_file_ = generated_random_file + f"_images_alpha_{alpha}_epoch_{epoch}.png"
image_path = os.path.join(save_dir, generated_random_file_)
save_image(generated_images, image_path)
else:
for alpha in alphas:
random_z = alpha*random_z_0 + (1.-alpha)*random_z_1
generated_images = model.decoder(random_z)
if model_outputs_logits:
generated_images = (torch.sigmoid(generated_images)).round().cpu()
else: # assuming input images are in the range of -1 and 1
generated_images = ((generated_images + 1) / 2).cpu()
generated_random_file_ = generated_random_file + f"_latent_alpha_{alpha}_epoch_{epoch}.png"
image_path = os.path.join(save_dir, generated_random_file_)
save_image(generated_images, image_path)
#Hyperparameters
batch_size = 64
lr = 3e-4
epochs = 20
save_interval = 3 #save model every save_interval epochs (None for not saving)
log_interval = 100 #print results every log_interval batches (in addition to every epoch (None for not printing)
seed = 1111
np.random.seed(seed)
torch.random.manual_seed(seed)
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
if __name__ == "__main__":
#Load Data
train_path = "binarized_mnist_train.amat"
val_path = "binarized_mnist_valid.amat"
test_path = "binarized_mnist_test.amat"
train_loader, val_loader, test_loader = load_data(train_path, test_path, val_path, batch_size=batch_size)
#Create Save Directory
save_parent_dir = os.curdir
model_folder = "VAE_binary_model_"
previous_folders = [f for f in os.listdir(save_parent_dir) if (model_folder in f and os.path.isdir(f))]
if not previous_folders:
save_dir = os.path.join(save_parent_dir, model_folder + f'{1:03d}')
else:
last_model = max(previous_folders)
i = int(last_model.replace(model_folder,"")) + 1
save_dir = os.path.join(save_parent_dir, model_folder + f'{i:03d}')
os.mkdir(save_dir)
#Create Model
model = VAE()
model = model.to(device)
optimizer = optim.Adam(model.parameters(), lr=lr)
loss_fn = Elbo_BCE()
train_samples, _ = next(iter(train_loader))
val_samples, _ = next(iter(val_loader))
random_z = torch.randn((batch_size, 100))
generate_samples(model, save_dir=save_dir, epoch=0, train_samples=train_samples, val_samples=val_samples, random_z=random_z)
print_model_summary(model, optimizer, save_dir=save_dir)
#Train Model
train_elbos, val_elbos, epoch_time = train(model, optimizer, train_loader, val_loader, loss_fn, epochs=epochs,
save_dir=save_dir, save_interval=save_interval, log_interval=log_interval,
train_samples=train_samples, val_samples=val_samples, random_z=random_z)
#generate num_samples random samples
# generate_random_samples(model, save_dir, epoch=epochs-1, num_samples=500)
#Report Results
test_elbo = epoch_eval(model, test_loader, loss_fn)
training_log_likelihood = estimate_data_likelihood(model, train_loader)
training_log_likelihood = sum(training_log_likelihood) / len(training_log_likelihood)
validation_log_likelihood = estimate_data_likelihood(model, val_loader)
validation_log_likelihood = sum(validation_log_likelihood) / len(validation_log_likelihood)
test_log_likelihood = estimate_data_likelihood(model, test_loader)
test_log_likelihood = sum(test_log_likelihood) / len(test_log_likelihood)
results_summary = f"Epoch: {epochs - 1}, Train Elbo: {train_elbos[-1]:.2f}, Validation Elbo: {val_elbos[-1]:.2f}, Test Elbo: {test_elbo:.2f}, " \
f"Training Log Likelihood: {training_log_likelihood:.2f}, Validation Log Likelihood: {validation_log_likelihood:.2f}, " \
f"Test Log Likelihood: {test_log_likelihood:.2f}, Epoch Time: {epoch_time:.2f}"
print(results_summary)
if save_interval is not None:
results_path = os.path.join(save_dir, 'results.txt')
with open(results_path, 'w') as file:
file.write(results_summary)