-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
137 lines (118 loc) · 4.72 KB
/
train.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
import torch
import torch.nn.functional as F
import numpy as np
from tqdm import tqdm
import os
# torch.autograd.set_detect_anomaly(True) #this line can have huge/enormous performance impact
# train the model
# training step
def train(model, train_loader, optimizer, epoch, device):
model.train()
# train_loss_averager = make_averager() # mantain a running average of the loss
# TRAIN
tqdm_iterator = tqdm(
enumerate(train_loader),
total=len(train_loader),
desc="",
leave=True,
)
len_tr_dl_ds = len(train_loader.dataset)
for batch_idx, (data, target) in tqdm_iterator:
data, target = data.to(device, non_blocking=True), target.to(
device, non_blocking=True
)
output = model(data)
loss = F.cross_entropy(output, target)
model.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
optimizer.zero_grad(set_to_none=True)
# train_loss_averager(loss.item())
tqdm_iterator.set_description(
f"Train Epoch: {epoch} [ {batch_idx * len(data)}/{len_tr_dl_ds} \tLoss: {loss.item():.6f}]")
tqdm_iterator.close()
if np.isnan(loss.item()):
print("Loss is nan")
raise Exception("Loss is nan")
exit()
# testing step
def test(model, test_loader, epoch, device):
model.eval()
test_loss = 0
correct = 0
tqdm_iterator = tqdm(
enumerate(test_loader),
total=len(test_loader),
desc="",
leave=True,
)
len_ts_dl_ds = len(test_loader.dataset)
with torch.no_grad():
for batch_idx, (data, target) in tqdm_iterator:
data, target = data.to(device, non_blocking=True), target.to(
device, non_blocking=True
)
output = model(data)
test_loss += F.cross_entropy(output, target).item() # sum up batch loss
# get the index of the max probability
pred = output.max(1, keepdim=True)[1] # equal to argmax
correct += pred.eq(target.view_as(pred)).cpu().sum().item()
tqdm_iterator.set_description(
f"Test Epoch: {epoch} [{batch_idx * len(data)}/{len_ts_dl_ds} \tLoss: {test_loss:.6f}, Accuracy: {correct}/{len_ts_dl_ds} ({100.0 * correct / len_ts_dl_ds}%)"
)
test_loss /= len(test_loader.dataset)
print(f"Validation Average loss: {test_loss:.6f}")
tqdm_iterator.close()
# show an histogram of the weights of the model
"""start = -1
stop = 1
bins = 30
for param in model.parameters():
if param.requires_grad:
hist = torch.histc(param.data, bins = bins, min = start, max = stop)
x = np.arange(start, stop, (stop-start)/bins)
plt.bar(x, hist.cpu(), align='center')
plt.ylabel('Frequency')
plt.show() """
return correct / len(test_loader.dataset), test_loss
def fit(
args,
model_intrinsic,
device,
train_dataloader,
test_dataloader,
optimizer,
val_global_accuracy,
):
best_acc, best_epoch = 0, 0
for epoch in range(1, args.num_epochs + 1):
train(model_intrinsic, train_dataloader, optimizer, epoch, device)
accuracy, val_avg_loss = test(model_intrinsic, test_dataloader, epoch, device)
print("Validation Accuracy: {}".format(accuracy))
# save information to file
with open(args.training_result_file, "a") as f:
f.write(f"\nEpoch: {epoch}")
f.write(f"\nValidation Average Loss: {val_avg_loss}")
f.write(f"\nValidation Accuracy: {accuracy}")
f.close()
# save the model if it reach validation accuracy > 90% of the actual chosen global accuracy
if accuracy > best_acc:
best_acc = accuracy
best_epoch = epoch
if best_acc >= val_global_accuracy:
saving_path = (
f"./{args.model_save_path}/{args.architecture}/{args.dataset}/"
)
if not os.path.exists(saving_path):
os.makedirs(saving_path)
if args.architecture == "fcn":
torch.save(
model_intrinsic.state_dict(),
f"{saving_path}{args.architecture}_h{args.hidden_dim}_id{args.intrinsic_dim}_lay{args.num_layers}_lr{args.learning_rate}_proj_{args.projection}_opt_{args.optimizer}.pt",
)
else:
torch.save(
model_intrinsic.state_dict(),
f"{saving_path}{args.architecture}_id{args.intrinsic_dim}_lr{args.learning_rate}_proj_{args.projection}_opt_{args.optimizer}.pt",
)
return epoch, best_epoch, accuracy, best_acc