-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_bbb.py
275 lines (211 loc) · 12.4 KB
/
train_bbb.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
# -*- coding: utf-8 -*-
"""
The file contains implementations of the functions used to train a Bayesian CNN model trained using Bayes by Backprop.
train_bnn - Function used to train a Bayesian Convolutional Neural Network using Bayes by Backprop.
"""
# Built-in/Generic Imports
import math
import time
# Library Imports
from torch.cuda import amp
from torch.nn import functional as F
from torch.utils.data import DataLoader
from torch.optim import SGD, lr_scheduler
from torch.utils.tensorboard import SummaryWriter
# Own Modules
from utils import *
from dataset import get_datasets
from model import BayesByBackpropClassifier
__author__ = ["Jacob Carse", "Andres Alvarez Olmo"]
__copyright__ = "Copyright 2022, Calibration"
__credits__ = ["Jacob Carse", "Andres Alvarez Olmo"]
__license__ = "MIT"
__version__ = "1.0.0"
__maintainer = ["Jacob Carse", "Andres Alvarez Olmo"]
__email__ = ["[email protected]", "[email protected]"]
__status__ = "Development"
def train_bnn(arguments: Namespace, device: torch.device) -> None:
"""
Function for training the Bayesian Convolutional Neural Network.
:param arguments: ArgumentParser Namespace object with arguments used for training.
:param device: PyTorch device that will be used for training.
"""
# Loads a TensorBoard Summary Writer.
if arguments.tensorboard_dir != "":
writer = SummaryWriter(os.path.join(arguments.tensorboard_dir, arguments.task, arguments.experiment))
# Loads the training and validation data.
train_data, val_data, _ = get_datasets(arguments)
# Creates the training data loader using the dataset objects.
training_data_loader = DataLoader(train_data, batch_size=arguments.batch_size,
shuffle=True, num_workers=arguments.data_workers,
pin_memory=False, drop_last=False)
# Creates the validation data loader using the dataset objects.
validation_data_loader = DataLoader(val_data, batch_size=arguments.batch_size * 2,
shuffle=False, num_workers=arguments.data_workers,
pin_memory=False, drop_last=False)
log(arguments, "Loader Datasets\n")
# Initialises the classifier model.
classifier = BayesByBackpropClassifier(arguments.efficient_net, train_data.num_class, device=device)
# Sets the classifier to training mode.
classifier.train()
# Moves the classifier to the selected device.
classifier.to(device)
# Initialises the optimiser used to optimise the parameters of the model.
optimiser = SGD(params=classifier.parameters(), lr=arguments.minimum_lr)
# Initialises the learning rate scheduler to adjust the learning rate during training.
scheduler = lr_scheduler.CyclicLR(optimiser, arguments.minimum_lr, arguments.maximum_lr, mode="triangular2")
# Initialises the gradient scaler used 16 bit precision.
if arguments.precision == 16 and device != torch.device("cpu"):
scaler = amp.GradScaler()
log(arguments, "Model Initialised")
# Declares the main logging variables for training.
start_time = time.time()
best_loss, best_epoch, total_batches = 1e10, 0, 0
batches_per_epoch = math.ceil(len(train_data) / arguments.batch_size)
batches_per_val = math.ceil(len(val_data) / arguments.batch_size * 2)
# The beginning of the main training loop.
for epoch in range(1, arguments.epochs + 1):
# Declares the logging variables for the epoch.
epoch_acc, epoch_loss, epoch_entropy, epoch_elbo, num_batches = 0., 0., 0., 0., 0
# Loops through the training data batches.
for images, labels in training_data_loader:
images = images.to(device)
labels = labels.to(device)
# Resets the gradients in the model.
optimiser.zero_grad()
# Performs training step with 16 bit precision.
if arguments.precision == 16 and device != torch.device("cpu"):
with amp.autocast():
# Samples the ELBO from the model and gets the average of the predictions.
kl_loss, predictions = classifier.sample_elbo(images, arguments.training_samples)
# Calculates the KL weight.
weight = 2 ** (batches_per_epoch - num_batches) / (2 ** batches_per_epoch - 1)
# Calculates the cross entropy loss and ELBO loss.
cross_entropy = F.cross_entropy(predictions, labels)
elbo = kl_loss * weight
# Combines the ELBO loss and Cross Entropy loss.
loss = cross_entropy + elbo
# Using the gradient scaler performs backward propagation.
scaler.scale(loss).backward()
# Update the weights of the model using the optimiser.
scaler.step(optimiser)
# Updates the scale factor of the gradient scaler.
scaler.update()
# Performs training step with 32 bit precision.
else:
# Samples the ELBO from the model and gets the average of the predictions.
kl_loss, predictions = classifier.sample_elbo(images, arguments.training_samples)
# Calculates the KL weight.
weight = 2 ** (batches_per_epoch - num_batches) / (2 ** batches_per_epoch - 1)
# Calculates the cross entropy loss and ELBO loss.
cross_entropy = F.cross_entropy(predictions, labels)
elbo = kl_loss * weight
# Combines the ELBO loss and Cross Entropy loss.
loss = cross_entropy + elbo
# Performs backward propagation.
loss.backward()
# Updates the weights of the model using the optimiser.
optimiser.step()
# Updates the learning rate scheduler.
scheduler.step()
# Calculates the accuracy of the batch.
batch_accuracy = (predictions.max(dim=1)[1] == labels).sum().double() / labels.shape[0]
# Adds the number of batches, losses and accuracy to the epoch sum.
num_batches += 1
epoch_loss += loss.item()
epoch_entropy += cross_entropy.item()
epoch_elbo += elbo.item()
epoch_acc += batch_accuracy
# Writes the batch loss and accuracy to TensorBoard logger.
if arguments.tensorboard_dir != "":
writer.add_scalar("Loss/batch", loss.item(), num_batches + total_batches)
writer.add_scalar("Cross Entropy/batch", cross_entropy.item(), num_batches + total_batches)
writer.add_scalar("ELBO/batch", elbo.item(), num_batches + total_batches)
writer.add_scalar("Accuracy/batch", batch_accuracy, num_batches + total_batches)
# Logs the details of the epoch progress.
if num_batches % arguments.log_interval == 0:
log(arguments, "Time: {}s\tTrain Epoch: {} [{}/{}] ({:.0f}%)\tLoss: {:.6f}\t"
"Cross Entropy: {:.4f}\tELBO: {:.4f}\tAccuracy: {:.6f}".format(
str(int(time.time() - start_time)).rjust(6, '0'), str(epoch).rjust(2, '0'),
str(num_batches * arguments.batch_size).rjust(len(str(len(train_data))), '0'),
len(train_data), 100. * num_batches / (len(train_data) / arguments.batch_size),
epoch_loss / num_batches, epoch_entropy / num_batches,
epoch_elbo / num_batches, epoch_acc / num_batches
))
# If the number of batches have been reached end epoch.
if num_batches == arguments.batches_per_epoch:
break
# Updates the total number of batches (used for debugging).
total_batches += num_batches
# Writes epoch loss and accuracy to TensorBoard.
if arguments.tensorboard_dir != "":
writer.add_scalar("Loss/train", epoch_loss, epoch)
writer.add_scalar("Cross Entropy/train", epoch_entropy, epoch)
writer.add_scalar("ELBO/train", epoch_elbo, epoch)
writer.add_scalar("Accuracy/train", epoch_acc, epoch)
# Declares the logging variables for validation.
val_acc, val_loss, val_entropy, val_elbo, val_batches = 0., 0., 0., 0., 0
# Performs the validation epoch with no gradient calculations.
with torch.no_grad():
# Loops through the validation data batches.
for images, labels, _ in validation_data_loader:
# Moves the images and labels to the selected device.
images = images.to(device)
labels = labels.to(device)
# Performs forward propagation using 16 bit precision.
if arguments.precision == 16 and device != torch.device("cpu"):
with amp.autocast():
# Samples the ELBO from the model and gets the average of the predictions.
kl_loss, predictions = classifier.sample_elbo(images, arguments.training_samples)
# Calculates the KL weight.
weight = 2 ** (batches_per_val - val_batches) / (2 ** batches_per_val - 1)
# Calculates the cross entropy loss and ELBO loss.
cross_entropy = F.cross_entropy(predictions, labels)
elbo = kl_loss * weight
# Combines the ELBO loss and Cross Entropy loss.
loss = cross_entropy + elbo
# Performs forward propagation using 32 bit precision.
else:
# Samples the ELBO from the model and gets the average of the predictions.
kl_loss, predictions = classifier.sample_elbo(images, arguments.training_samples)
# Calculates the KL weight.
weight = 2 ** (batches_per_val - val_batches) / (2 ** batches_per_val - 1)
# Calculates the cross entropy loss and ELBO loss.
cross_entropy = F.cross_entropy(predictions, labels)
elbo = kl_loss * weight
# Combines the ELBO loss and Cross Entropy loss.
loss = cross_entropy + elbo
# Calculates the accuracy of the batch.
batch_accuracy = (predictions.max(dim=1)[1] == labels).sum().double() / labels.shape[0]
# Adds the number of batches, loss and accuracy to validation sum.
val_batches += 1
val_loss += loss.item()
val_entropy += cross_entropy.item()
val_elbo += elbo.item()
val_acc += batch_accuracy.item()
# If the number of batches have been reached end validation.
if val_batches == arguments.batches_per_epoch:
break
# Writes validation loss and accuracy to TensorBoard
if arguments.tensorboard_dir != "":
writer.add_scalar("Loss/val", val_loss / val_batches, epoch)
writer.add_scalar("Cross Entropy/val", val_entropy / val_batches, epoch)
writer.add_scalar("ELBO/val", val_elbo / val_batches, epoch)
writer.add_scalar("Accuracy/val", val_acc / val_batches, epoch)
# Logs the details of the training epoch.
log(arguments, "\nEpoch: {}\tTraining Loss: {:.6f}\tTraining Cross Entropy: {:.6f}\t"
"Training ELBO: {:.6f}\tTraining Accuracy: {:.6f}\n"
"Validation Loss: {:.6f}\tValidation Cross Entropy: {:.6f}\t"
"Validation ELBO: {:.6f}\tValidation Accuracy: {:.6f}\n".format(
epoch, epoch_loss / num_batches, epoch_entropy / num_batches, epoch_elbo / num_batches,
epoch_acc / num_batches, val_loss / val_batches, val_entropy / val_batches, val_elbo / val_batches,
val_acc / val_batches
))
# If the current epoch has the best validation loss then save the model with the prefix best.
if val_loss / val_batches < best_loss:
best_loss = val_loss / val_batches
best_epoch = epoch
classifier.save_model(arguments.model_dir, arguments.experiment)
# Logs the final training information.
log(arguments,
f"\nTraining Finished with best loss of {best_loss} at epoch {best_epoch} in {int(time.time() - start_time)}s.")