-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrainer.py
219 lines (174 loc) · 10.3 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
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
import sys
import pickle
from tqdm import tqdm
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import torch
import os
import copy
import random
import shutil
import datetime
import matplotlib.pyplot as plt
from path import Path
from torch.utils.tensorboard import SummaryWriter
from models.losses import SoftTargetKDLoss
import helpers
"""
Trainer class to train and evaluate the model
"""
class Trainer:
def __init__(self, opts, net, optimizer, scheduler, train_dataloader, test_dataloader):
self.opts = opts
self.net = net
self.optimizer = optimizer
self.scheduler = scheduler
self.train_loader = train_dataloader
self.test_loader = test_dataloader
# Log
if not self.opts['eval']:
self.best_accuracy_orig, self.best_accuracy_cape, self.best_epoch = 0, 0, 0
self.train_losses, self.test_losses = [], []
self.train_accuracies_orig, self.test_accuracies_orig = [], []
self.train_accuracies_cape, self.test_accuracies_cape = [], []
self.writer = SummaryWriter(log_dir=self.opts['experiment_dir'])
# Loss
self.kl_loss = SoftTargetKDLoss(T=self.opts['T_kld']).to(self.opts['device'])
self.classification_loss = nn.CrossEntropyLoss().to(self.opts['device'])
@staticmethod
def requires_grad(model, flag=True):
for p in model.parameters():
p.requires_grad = flag
def run(self):
for epoch_no in range(self.opts['num_epochs']):
disp_epoch_no = epoch_no + 1
print('Learning Rate: {}'.format(self.optimizer.param_groups[0]['lr']))
self.writer.add_scalar('Learning rate', self.optimizer.param_groups[0]['lr'], epoch_no)
self.train(epoch_no)
train_results = f"{datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')} Epoch {disp_epoch_no}/{self.opts['num_epochs']} Train Loss: {self.train_losses[-1]}, Train Accuracy [orig]: {self.train_accuracies_orig[-1]} Train Accuracy [cape]: {self.train_accuracies_cape[-1]}"
print(train_results)
self.test(epoch_no)
test_results = f"{datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')} Epoch {disp_epoch_no}/{self.opts['num_epochs']} Test Loss: {self.test_losses[-1]}, Test Accuracy [orig]: {self.test_accuracies_orig[-1]} Test Accuracy [cape]: {self.test_accuracies_cape[-1]}"
print(test_results)
# break
# scheduler
if self.scheduler is not None:
self.scheduler.step()
if np.mod(disp_epoch_no, self.opts['save_interval']) == 0:
if isinstance(self.net, nn.DataParallel):
net_state_dict = self.net.module.state_dict()
else:
net_state_dict = self.net.state_dict()
torch.save(net_state_dict, os.path.join(self.opts['experiment_dir'], 'net_e{}.ckpt'.format(disp_epoch_no)))
if self.best_accuracy_orig <= self.test_accuracies_orig[-1] and self.best_accuracy_cape <= self.test_accuracies_cape[-1]:
self.best_epoch = disp_epoch_no
log = f"Improve accuracy from [orig]: {self.best_accuracy_orig} to {self.test_accuracies_orig[-1]} and [cape]: {self.best_accuracy_cape} to {self.test_accuracies_cape[-1]}"
print(log)
self.best_accuracy_orig = self.test_accuracies_orig[-1]
self.best_accuracy_cape = self.test_accuracies_cape[-1]
if isinstance(self.net, nn.DataParallel):
net_state_dict = self.net.module.state_dict()
else:
net_state_dict = self.net.state_dict()
torch.save(net_state_dict, os.path.join(self.opts['experiment_dir'], f'best.pth'))
print(f"Best Accuracy [orig]: {self.best_accuracy_orig} and [cape]: {self.best_accuracy_cape} in epoch {self.best_epoch}")
self.writer.close()
def train(self, epoch):
t = tqdm(enumerate(self.train_loader, 0), total=len(self.train_loader),
smoothing=0.9, position=0, leave=True,
desc="Train: Epoch: "+str(epoch+1)+"/"+str(self.opts['num_epochs']))
self.net.train()
if self.opts['settings'] == 'PF':
# here batch normalization layers are now fully freezed by putting them in eval mode
if isinstance(self.net, nn.DataParallel):
self.net.module.net.eval()
else:
self.net.net.eval()
num_total, acc_orig, acc_cape = 0, 0, 0
running_total_loss, running_kld_loss, running_cls_loss = 0, 0, 0
for i, (images, targets) in t:
images = images.to(self.opts['device'])
targets = targets.type(torch.LongTensor).to(self.opts['device'])
self.optimizer.zero_grad()
outputs = self.net(images)
class_loss = self.opts['loss_alpha'] * self.classification_loss(outputs['orig']['outcome'], targets)
kld_loss = self.opts['loss_beta'] * self.kl_loss(outputs['cape']['outcome_soft'], outputs['orig']['outcome'].detach())
loss = class_loss + kld_loss
loss.backward()
self.optimizer.step()
_, pred_orig = torch.max(outputs['orig']['outcome'].softmax(dim=1), 1)
acc_orig += (pred_orig == targets).sum().item()
_, pred_cape = torch.max(outputs['cape']['outcome'], 1)
acc_cape += (pred_cape == targets).sum().item()
num_total += targets.size(0)
running_total_loss += loss.item()
running_cls_loss += class_loss.item()
running_kld_loss += kld_loss.item()
t.set_postfix_str(f'class_loss: {class_loss.item():.4} kld_loss: {kld_loss.item():.4}')
self.train_accuracies_orig.append(100*acc_orig / num_total)
self.train_accuracies_cape.append(100*acc_cape / num_total)
self.train_losses.append(running_total_loss / len(self.train_loader))
# save summary
self.writer.add_scalar('Loss/train', running_total_loss / len(self.train_loader), epoch)
self.writer.add_scalar('KLD Loss/train', running_kld_loss / len(self.train_loader), epoch)
self.writer.add_scalar('CLS Loss/train', running_cls_loss / len(self.train_loader), epoch)
self.writer.add_scalar('Accuracy [orig]/train', 100*acc_orig / num_total, epoch)
self.writer.add_scalar('Accuracy [cape]/train', 100*acc_cape / num_total, epoch)
def test(self, epoch):
t = tqdm(enumerate(self.test_loader, 0), total=len(self.test_loader),
smoothing=0.9, position=0, leave=True,
desc="Test: Epoch: "+str(epoch+1)+"/"+str(self.opts['num_epochs']))
self.net.eval()
num_total, acc_orig, acc_cape = 0, 0, 0
running_total_loss, running_kld_loss, running_cls_loss = 0, 0, 0
with torch.no_grad():
for i, (images, targets) in t:
images = images.to(self.opts['device'])
targets = targets.type(torch.LongTensor).to(self.opts['device'])
outputs = self.net(images)
class_loss = self.opts['loss_alpha'] * self.classification_loss(outputs['orig']['outcome'], targets)
kld_loss = self.opts['loss_beta'] * self.kl_loss(outputs['cape']['outcome_soft'], outputs['orig']['outcome'].detach())
loss = class_loss + kld_loss
_, pred_orig = torch.max(outputs['orig']['outcome'].softmax(dim=1), 1)
acc_orig += (pred_orig == targets).sum().item()
_, pred_cape = torch.max(outputs['cape']['outcome'], 1)
acc_cape += (pred_cape == targets).sum().item()
num_total += targets.size(0)
running_total_loss += loss.item()
running_cls_loss += class_loss.item()
running_kld_loss += kld_loss.item()
# break
t.set_postfix_str(f'class_loss: {class_loss.item():.4} kld_loss: {kld_loss.item():.4}')
self.test_accuracies_orig.append(100*acc_orig / num_total)
self.test_accuracies_cape.append(100*acc_cape / num_total)
self.test_losses.append(running_total_loss / len(self.test_loader))
# save summary
self.writer.add_scalar('Loss/test', running_total_loss / len(self.test_loader), epoch)
self.writer.add_scalar('KLD Loss/test', running_kld_loss / len(self.test_loader), epoch)
self.writer.add_scalar('CLS Loss/test', running_cls_loss / len(self.test_loader), epoch)
self.writer.add_scalar('Accuracy [orig]/test', 100*acc_orig / num_total, epoch)
self.writer.add_scalar('Accuracy [cape]/test', 100*acc_cape / num_total, epoch)
def eval(self):
t = tqdm(enumerate(self.test_loader, 0), total=len(self.test_loader),
smoothing=0.9, position=0, leave=True,
desc="Evaluation: ")
self.net.eval()
with torch.no_grad():
num_acc_orig, num_acc_camp = 0, 0
num_total = 0
vanilla_sum, campe_sum = 0, 0
for i, (images, targets) in t:
images = images.to(self.opts['device'])
num_total+=images.size(0)
targets = targets.float().to(self.opts['device'])
outputs = self.net(images)
pred_vals, pred_orig = torch.max(outputs['orig']['outcome'].softmax(dim=1), 1)
vanilla_sum+=pred_vals.sum().item()
num_acc_orig += (pred_orig == targets).sum().item()
pred_vals, pred = torch.max(outputs['cape']['outcome'], 1)
campe_sum+=pred_vals.sum().item()
num_acc_camp += (pred == targets).sum().item()
print(f'Net=> Acc orig: {round(num_acc_orig*100/num_total, 3)} Acc camp: {round(num_acc_camp*100/num_total, 3)}')
# empirical mean of prediction confidence
print(f'Net=> Avg. conf orig: {round(vanilla_sum*100/num_total, 3)} Avg. conf. camp: {round(campe_sum*100/num_total, 3)}')