/
trainer.py
181 lines (141 loc) · 7.04 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
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as fn
from tensorboardX import SummaryWriter
from torch_geometric.data import DataLoader
from tqdm import tqdm
from args import make_args
from data.dataset3 import SkeletonDataset
from models.net import DualGraphEncoder
from optimizer import get_std_opt
class GCNTrainer(object):
def __init__(self, model, train_loader, val_loader, adj, optimizer, loss_fn, log_dir):
self.model = model
self.train_loader = train_loader
self.val_loader = val_loader
self.log_dir = log_dir
self.adj = adj
self.optimizer = optimizer
self.num_classes = 60
self.device = torch.device('cuda:0')
self.loss_fn = nn.CrossEntropyLoss().to(self.device)
self.model = self.model.to(self.device)
self.adj = self.adj.to(self.device)
if self.log_dir is not None:
self.writer = SummaryWriter(log_dir)
def train(self, n_epochs):
best_acc = 0
i_acc = 0
for epoch in range(n_epochs):
self.model.train(True)
# plot learning rate
lr = self.optimizer.state_dict()['param_groups'][0]['lr']
self.writer.add_scalar('params/lr', lr, epoch)
for i, batch in tqdm(enumerate(train_loader), total=len(train_loader), desc="Epoch {}".format(epoch)):
batch = batch.to(self.device)
#self.optimizer.zero_grad()
output = self.model(batch.x, adj=self.adj, bi=batch.batch)
target = batch.y # - 1
# one_hot = fn.one_hot(target.long(), num_classes = 60)
# loss = fn.cross_entropy(output, one_hot)l
loss = fn.cross_entropy(output, target.long())
# loss_value = loss.cpu().item()
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
self.writer.add_scalar('train/train_loss', loss, i_acc + i + 1)
pred = torch.max(output, 1)[1]
results = pred == target
correct_points = torch.sum(results.long())
acc = correct_points.float() / results.size()[0]
self.writer.add_scalar('train/train_overall_acc', acc, i_acc + i + 1)
for name, param in self.model.named_parameters():
if param.requires_grad and param.grad is not None:
self.writer.add_scalar('gradients/' + name, param.grad.norm(2).item(), i + 1)
# log_str = 'epoch %d, step %d: train_loss %.3f; train_acc %.3f' % (epoch + 1, i + 1, loss, acc)
# log_str = 'epoch %d, step %d: train_loss %.3f' % (epoch + 1, i + 1, loss)
# if (i + 1) % 1 == 0:
# print(log_str)
i_acc += i
# evaluation
with torch.no_grad():
# loss, val_overall_acc, val_mean_class_acc = self.update_validation_accuracy()
loss, val_overall_acc = self.update_validation_accuracy()
# self.writer.add_scalar('val/val_mean_class_acc', val_mean_class_acc, epoch+1)
self.writer.add_scalar('val/val_overall_acc', val_overall_acc, epoch + 1)
self.writer.add_scalar('val/val_loss', loss, epoch + 1)
# save best model
if val_overall_acc > best_acc:
best_acc = val_overall_acc
# self.model.save(self.log_dir, epoch)
torch.save(self.model.state_dict(),
os.path.join(self.log_dir,
"best_model.pth"))
# adjust learning rate manually
if epoch > 0 and (epoch + 1) % 10 == 0:
for param_group in self.optimizer.param_groups:
param_group['lr'] = param_group['lr'] * 0.5
# export scalar data to JSON for external processing
self.writer.export_scalars_to_json(self.log_dir + "/all_scalars.json")
self.writer.close()
def update_validation_accuracy(self):
all_correct_points = 0
all_points = 0
wrong_class = np.zeros(self.num_classes)
samples_class = np.zeros(self.num_classes)
all_loss = 0
self.model.eval()
total_time = 0.0
total_print_time = 0.0
all_target = []
all_pred = []
for i, batch in enumerate(self.val_loader, 0):
batch = batch.to(self.device)
output = self.model(batch.x, adj=self.adj, bi=batch.batch)
target = batch.y
pred = torch.max(output, 1)[1]
all_loss += self.loss_fn(output, target).cpu().data.numpy()
results = pred == target
for i in range(results.size()[0]):
if not bool(results[i].cpu().data.numpy()):
wrong_class[target.cpu().data.numpy().astype('int')[i]] += 1
samples_class[target.cpu().data.numpy().astype('int')[i]] += 1
correct_points = torch.sum(results.long())
all_correct_points += correct_points
all_points += results.size()[0]
print('Total # of test models: ', all_points)
# val_mean_class_acc = np.mean((samples_class - wrong_class) / samples_class)
acc = all_correct_points.float() / all_points
val_overall_acc = acc.cpu().data.numpy()
loss = all_loss / len(self.val_loader)
# print ('val mean class acc. : ', val_mean_class_acc)
print('val overall acc. : ', val_overall_acc)
print('val loss : ', loss)
self.model.train()
# return loss, val_overall_acc, val_mean_class_acc
return loss, val_overall_acc
if __name__ == '__main__':
args = make_args()
log_dir = '/home/dusko/Documents/projects/APBGCN/log'
train_dataset = SkeletonDataset(root="/home/dusko/Documents/projects/APBGCN", name='ntu_60',
use_motion_vector=False,
benchmark='xsub', sample='train')
valid_dataset = SkeletonDataset(root="/home/dusko/Documents/projects/APBGCN", name='ntu_60',
use_motion_vector=False,
benchmark='xsub', sample='val')
train_loader = DataLoader(train_dataset.data, batch_size=args.batch_size, shuffle=True)
valid_loader = DataLoader(valid_dataset.data, batch_size=args.batch_size, shuffle=True)
model = DualGraphEncoder(in_channels=6,
hidden_channels=128,
out_channels=128,
num_layers=8,
num_heads=8,
linear_temporal=True,
sequential=False)
# optimizer = optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.weight_decay, betas=(0.9, 0.98))
noam_opt = get_std_opt(model, args)
trainer = GCNTrainer(model, train_loader, valid_loader,
train_dataset.skeleton_, noam_opt.optimizer, nn.CrossEntropyLoss(), log_dir)
trainer.train(args.epoch_num)