-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_model_contrastive.py
197 lines (156 loc) · 8.2 KB
/
train_model_contrastive.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
import torch
from utils.evaluate import evaluate_model_contrastive
from tqdm import tqdm
import os
from time import gmtime, strftime
from shutil import copyfile
import numpy as np
from models.STGCN import get_normalized_adj
import yaml
from utils.ModelMonitor import ModelMonitoring
def train(moco_encoder, linear, loader_train, optimizer, scheduler, encoder_loss, classifier_loss, wandb,epochs=200,device="cuda:2", test=False, loader_test=None, log_model=20, output_dir=None, adj=None, config_file=None, sklt=False):
e = 0
log_dir = strftime("%d-%m-%y %H:%M:%S", gmtime())
output_dir = os.path.join(output_dir,log_dir)
if not os.path.exists(output_dir):
print(f"{output_dir} directory created")
os.makedirs(output_dir)
os.makedirs(os.path.join(output_dir,"encoder"))
os.makedirs(os.path.join(output_dir,"linear"))
copyfile(config_file, os.path.join(output_dir,"params.yaml"))
moco_encoder.train()
linear.train()
with open(adj, 'rb') as f:
A = np.load(f)
if A.sum() != 51**2:
A = A + np.identity(51)
A_hat = torch.Tensor(get_normalized_adj(A)).to(device)
with open(config_file) as f:
conf = yaml.safe_load(f)
if conf["training"]["audio_only"]:
num_nodes = conf["dataset"]["n_mels"]
num_feat_in = 1
adj = np.ones((num_nodes,num_nodes))- np.identity(num_nodes)
A_hat = torch.Tensor(get_normalized_adj(adj)).to(device)
inspector = ModelMonitoring(patience=conf["training"]["patience"])
optimizer_encoder, optimizer_decoder = optimizer[0], optimizer[1]
scheduler_encoder, scheduler_decoder = scheduler[0], scheduler[1]
n_classes = conf["dataset"]["classes"]
for e in range(epochs):
samples = 0.
batch_count = 0
cumulative_loss = 0.
cumulative_contr_loss = 0.
cumulative_ce_loss = 0.
cumulative_accuracy = 0.
train_label_pred = [0 for k in range(n_classes)]
train_label_pred_count = [0 for k in range(n_classes)]
train_label_count = [0 for k in range(n_classes)]
print(f"Epoch - {e}")
for batch_idx, batch in enumerate(tqdm(loader_train)):
if len(batch) ==3:
targets, ld_1, ld_2 = batch[0].to(device),batch[1].to(device), batch[2].to(device)
else:
targets, ld_1, ld_2, ad_1, ad_2 = batch[0].to(device), batch[1].to(device), batch[2].to(device), batch[3].to(device),batch[4].to(device)
targets = targets.long()
#print(ld_1.shape)
#print(ld_1)
# Forward pass
#contr_feat, contr_tar, video_features = moco_encoder(ld_1,ld_2,targets, train=True)
if len(batch) ==3:
q1, vf_q1 = moco_encoder(A_hat, ld_1)
q2, vf_q2 = moco_encoder(A_hat, ld_2)
else:
q1, vf_q1 = moco_encoder(ld_1, ad_1)
q2, vf_q2 = moco_encoder(ld_2, ad_2)
contr_feat = torch.cat((q1.unsqueeze(1),q2.unsqueeze(1)),1)
if conf["training"]["unsupervised"]:
contr_loss = encoder_loss(contr_feat)
else:
contr_loss = encoder_loss(contr_feat, targets)
video_feat = torch.cat((vf_q1.detach(),vf_q2.detach()),0)
logits = linear(video_feat)
ce_loss = classifier_loss(logits, torch.cat((targets,targets),0).long())
targets = torch.cat((targets,targets))
loss = contr_loss + ce_loss
optimizer_encoder.zero_grad()
optimizer_decoder.zero_grad()
contr_loss.backward()
ce_loss.backward()
optimizer_encoder.step()
optimizer_decoder.step()
batch_size = ld_1.shape[0]
samples += batch_size*2
batch_count +=1
cumulative_loss += loss.item()
cumulative_contr_loss += contr_loss.item() # Note: the .item() is needed to extract scalars from tensors
cumulative_ce_loss += ce_loss.item() # Note: the .item() is needed to extract scalars from tensors
_, predicted = logits.max(1)
cumulative_accuracy += predicted.eq(targets).sum().item()
for i in range(predicted.shape[0]):
if predicted[i] == targets[i]:
train_label_pred[predicted[i]] +=1
train_label_count[targets[i]] +=1
train_label_pred_count[predicted[i]] += 1
scheduler_encoder.step()
scheduler_decoder.step()
final_loss = cumulative_loss/batch_count
final_contr_loss = cumulative_contr_loss/batch_count
final_ce_loss = cumulative_ce_loss/batch_count
accuracy = cumulative_accuracy/samples*100
if e % log_model == 0:
filename = os.path.join(output_dir,"encoder","encoder_epoch_"+str(e)+".pth")
torch.save({
'epoch': e,
'model_state_dict': moco_encoder.state_dict(),
'optimizer_state_dict': optimizer_encoder.state_dict(),
'loss': final_loss,
}, filename)
filename = os.path.join(output_dir,"linear","linear_epoch_"+str(e)+".pth")
torch.save({
'epoch': e,
'model_state_dict': linear.state_dict(),
'optimizer_state_dict': optimizer_decoder.state_dict(),
'loss': final_loss,
}, filename)
# test performance over the test set
if test:
test_loss, test_contr_loss, test_ce_loss, test_accuracy, label_pred_correct, label_pred_count, label_tot_count = evaluate_model_contrastive(moco_encoder,linear, loader_test, encoder_loss, classifier_loss,A_hat, device=device, unsupervised=conf["training"]["unsupervised"], n_classes=n_classes)
print('\t Test loss {:.5f}, Test_contr_loss {:.5f}, Test_ce_loss {:.5f}, Test accuracy {:.2f}'.format(test_loss, test_contr_loss, test_ce_loss,test_accuracy))
if wandb is not None:
wandb.log({"Test_Accuracy": test_accuracy , "Test_Contrastive Loss": test_contr_loss,
"Test_Cross Entropy Loss": test_ce_loss, "Test_Total Loss": test_loss})
correct = label_pred_correct/label_tot_count
if n_classes==7:
label = ["Neutral","Anger","Disgust","Fear","Happiness","Sadness","Surprise"]
else:
label = ["neutral", "calm", "happy","sad", "angry", "fearful", "disgust", "surprised"]
for i in range(len(label_pred_count)):
wandb.log({"Test_label_percentage_"+str(label[i]): correct[i] })
inspector(test_accuracy)
print(f"BEST SCORE {inspector.best_score} count {inspector.counter}/{inspector.patience}")
print('\t Training loss {:.5f}, Train_contr_loss {:.5f}, Train_ce_loss {:.5f}, Training accuracy {:.2f}'.format(final_loss, final_contr_loss, final_ce_loss, accuracy))
if wandb is not None:
wandb.log({"Accuracy": accuracy, "Contrastive Loss": final_contr_loss,
"Cross Entropy Loss": final_ce_loss, "Total Loss": final_loss})
correct_train = np.array(train_label_pred)/np.array(train_label_count)
for i in range(len(label_pred_count)):
wandb.log({"Train_label_percentage_"+str(label[i]): correct_train[i] })
if inspector.stopped:
print(f"BEST SCORE {inspector.best_score}")
filename = os.path.join(output_dir,"encoder","encoder_epoch_"+str(e)+".pth")
torch.save({
'epoch': e,
'model_state_dict': moco_encoder.state_dict(),
'optimizer_state_dict': optimizer_encoder.state_dict(),
'loss': final_loss,
}, filename)
filename = os.path.join(output_dir,"linear","linear_epoch_"+str(e)+".pth")
torch.save({
'epoch': e,
'model_state_dict': linear.state_dict(),
'optimizer_state_dict': optimizer_decoder.state_dict(),
'loss': final_loss,
}, filename)
break
#