-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
192 lines (168 loc) · 5.93 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
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
from __future__ import print_function
import argparse
import os
import shutil
import time
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.backends.cudnn as cudnn
import torch.optim
#import torch.utils.data
import torch.nn.functional as F
import torchvision.transforms as transforms
import torchvision.datasets as datasets
#from torchsummary import summary
import torchvision.models as models
# from models import *
from collections import OrderedDict
from torch.autograd import Variable
# import scipy as sp
from scipy import signal
from tqdm import tqdm
from scipy.stats import pearsonr
from sklearn.metrics import mean_absolute_error
from sklearn.metrics import mean_squared_error
from utils.utils import clip_gradient
import utils.utils as utils
from utils.exp_utils import pearson
from EvaluationMetrics.ICC import compute_icc
from EvaluationMetrics.CCC_score import CCC_score
from utils.utils import Normalize
from utils.utils import calc_scores
import logging
# import models.resnet as ResNet
#import utils
import matplotlib.pyplot as plt
import numpy as np
# import cv2
import sys
import math
from losses.CCC import CCC
import wandb
learning_rate_decay_start = 15 # 50
learning_rate_decay_every = 2 # 5
learning_rate_decay_rate = 0.8 # 0.9
total_epoch = 30
lr = 0.0001
scaler = torch.cuda.amp.GradScaler()
def train(train_loader, audio_model, visual_model, audiovisual_model, criterion, optimizer, epoch, cam):
print('\nEpoch: %d' % epoch)
global Train_acc
wandb.watch(audiovisual_model, log_freq=100)
wandb.watch(cam, log_freq=100)
# switch to train mode
audiovisual_model.train()
audio_model.eval()
visual_model.eval()
cam.train()
train_loss = 0
correct = 0
total = 0
running_loss = 0
running_accuracy = 0
out = []
tar = []
if epoch > learning_rate_decay_start and learning_rate_decay_start >= 0:
frac = (epoch - learning_rate_decay_start) // learning_rate_decay_every
decay_factor = learning_rate_decay_rate ** frac
current_lr = lr * decay_factor
utils.set_lr(optimizer, current_lr) # set the decayed rate
else:
current_lr = lr
print('learning_rate: %s' % str(current_lr))
logging.info("Learning rate")
logging.info(current_lr)
#torch.cuda.synchronize()
#t1 = time.time()
for batch_idx, (visual_data, audiodata, labels) in tqdm(enumerate(train_loader),
total=len(train_loader), position=0, leave=True):
#if(batch_idx > 2):#int(65844/64)):
# break
#torch.cuda.synchronize()
#t2 = time.time()
#print('data loading time', t2-t1)
optimizer.zero_grad()
audiodata = audiodata.cuda()
visualdata = visual_data.cuda()#permute(0,4,1,2,3).cuda()
#visuallabel = visuallabel.squeeze(1).type(torch.FloatTensor).cuda()
#print("training started")
#torch.cuda.synchronize()
#t3 = time.time()
with torch.cuda.amp.autocast():
with torch.no_grad():
b, c, seq_t, subseq_t, h, w = visualdata.size()
#visualdata = visual_data.view(b, c, -1, sub_seq_len, h, w)
visual_feats = []
aud_feats = []
#vis_data = visualdata.view(b*visualdata.shape[2], c, subseq_t ,h , w)
#visualfeatures, _ = visual_model(vis_data)
#visual_feat = visualfeatures.view(b, -1, visualfeatures.shape[1])
#print(visfin_test.shape)
#print(visual_feat.shape)
for i in range(visualdata.shape[0]):
vis_dat = visualdata[i, :, :, :,:,:].transpose(0,1)
visualfeat, _ = visual_model(vis_dat)
#visual_feat = visualfeat.view(b, -1, visualfeat.shape[1])
visual_feats.append(visualfeat)
#aud_data = audiodata.view(audiodata.shape[0]*audiodata.shape[1], audiodata.shape[2], audiodata.shape[3]).unsqueeze(1)
aud_data = audiodata[i,:,:,:].unsqueeze(1)
aud_feat, audio_out = audio_model(aud_data)
#print(aud_feat.shape)
#audio_feat = aud_feat.view(b, -1, aud_feat.shape[1])
aud_feats.append(aud_feat.squeeze(3))
#print(audio_feat.shape)
visual_feat = torch.stack(visual_feats).squeeze(3).squeeze(3).squeeze(3)#.transpose(1,2)
#visual_feat = visual_feat.view(visual_feat.shape[0]*visual_feat.shape[1], -1)
#print(visual_feat.shape)
#torch.cuda.synchronize()
#t4 = time.time()
#print('visual feature extraction time', t4-t3)
#torch.cuda.synchronize()
#t5 = time.time()
#aud_feats = []
#print(audiodata.shape)
#for i in range(audiodata.shape[0]):
# aud_data = audiodata[i,:,:,:].unsqueeze(1)
# audio_feat, audio_out = audio_model(aud_data)
# aud_feats.append(audio_feat.squeeze(3))
audio_feat = torch.stack(aud_feats).squeeze(3)#.transpose(1,2)
#print(audio_feat.shape)
#print(visual_feat.shape)
#torch.cuda.synchronize()
#t6 = time.time()
#print('audio feature extraction time', t6-t5)
#audio_feat = audio_feat.squeeze(3)#.transpose(1,2)
#visual_feat = torch.max(visual_feat, dim = 2)[0]#.squeeze(2).squeeze(2)
#audio_feat = torch.max(audio_feat, dim = 2)[0]#.squeeze(2).squeeze(2)
#print("features extracted")
audio_feat_norm = F.normalize(audio_feat, p=2, dim=2, eps=1e-12)
visual_feat_norm = F.normalize(visual_feat, p=2, dim=2, eps=1e-12)
audio_attfeat, visual_attfeat = cam(audio_feat_norm, visual_feat_norm)
#with torch.no_grad():
audiovisual_outs = audiovisual_model(audio_attfeat, visual_attfeat)
outputs = audiovisual_outs.view(-1, audiovisual_outs.shape[0]*audiovisual_outs.shape[1])
targets = labels.view(-1, labels.shape[0]).cuda()
loss = criterion(outputs, targets)
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
out = np.concatenate([out, outputs.squeeze(0).detach().cpu().numpy()])
tar = np.concatenate([tar, targets.squeeze(0).detach().cpu().numpy()])
if torch.isnan(loss):
print(outputs)
print(targets)
print(loss)
sys.exit()
if batch_idx % 10 == 0:
wandb.log({"train_loss": loss})
pred, tar = Normalize(out, tar)
if (len(tar) > 1):
train_acc = CCC_score(tar, pred)
else:
train_acc = 0
print("Train Accuracy")
wandb.log({"train_acc": train_acc})
print(train_acc)
#xcorr_weights = 0
return (loss), (train_acc)