-
Notifications
You must be signed in to change notification settings - Fork 31
/
attention_visualization.py
149 lines (124 loc) · 4.83 KB
/
attention_visualization.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
from __future__ import print_function
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0" # set GPU ID
import torch.nn as nn
from torch.autograd import Variable
import matplotlib.pyplot as plt
from dataloader import *
import random
import models_fusion
import time
from sklearn.metrics import accuracy_score
import torch.nn.functional as F
import imageio
import cv2
from sklearn.preprocessing import MinMaxScaler
import warnings
warnings.filterwarnings("ignore")
scaler = MinMaxScaler(copy=True, feature_range=(0, 1))
def video_frame_sample(frame_interval, video_length, sample_num):
num = []
for l in range(video_length):
for i in range(sample_num):
num.append(int(l * frame_interval + (i * 1.0 / sample_num) * frame_interval))
return num
def normlize(x, min = 0, max = 255):
num, row, col = x.shape
for i in range(num):
xi = x[i, :, :]
xi = max *(xi - np.min(xi))/(np.max(xi) - np.min(xi))
x[i, :, :] = xi
return x
def create_heatmap(im_map, im_cloud, kernel_size=(5,5),colormap=cv2.COLORMAP_JET,a1=0.5,a2=0.3):
print(np.max(im_cloud))
im_cloud[:, :, 1] = 0
im_cloud[:, :, 2] = 0
return (a1*im_map + a2*im_cloud).astype(np.uint8)
# features, labels, and testing set list
dir_video = "data/visual_feature.h5"
dir_audio = "data/audio_feature.h5"
dir_labels = "data/labels.h5"
dir_order_test = "data/test_order.h5"
# access to original videos for extracting video frames
raw_video_dir = "data/AVE" # videos in AVE dataset
lis = os.listdir(raw_video_dir)
f = open("data/Annotations.txt", 'r')
dataset = f.readlines()
print("The dataset contains %d samples" % (len(dataset)))
f.close()
len_data = len(dataset)
with h5py.File(dir_order_test, 'r') as hf:
test_order = hf['order'][:]
# pre-trained models
att_model = torch.load('model/AV_att.pt')
att_layer = att_model._modules.get('affine_h') # extract attention maps from the layer
# load testing set
AVEData = AVEDataset(video_dir=dir_video, audio_dir=dir_audio, label_dir=dir_labels,
order_dir=dir_order_test, batch_size=402)
nb_batch = AVEData.__len__()
print(nb_batch)
audio_inputs, video_inputs, labels = AVEData.get_batch(0)
audio_inputs = Variable(audio_inputs.cuda(), requires_grad=False)
video_inputs = Variable(video_inputs.cuda(), requires_grad=False)
labels = labels.numpy()
# generate attention maps
att_map = torch.zeros((4020, 49, 1))
def fun(m, i, o): att_map.copy_(o.data)
map = att_layer.register_forward_hook(fun)
h_x = att_model(audio_inputs, video_inputs)
map.remove()
z_t = Variable(att_map.squeeze( 2 ))
alpha_t = F.softmax( z_t, dim = -1 ).view( z_t.size( 0 ), -1, z_t.size( 1 ) )
att_weight = alpha_t.view(402, 10, 7, 7).cpu().data.numpy() # attention maps of all testing samples
c = 0
t = 10
sample_num = 16 # 16 frames for 1-sec video segment
extract_frames = np.zeros((160, 224, 224, 3)) # 160 224x224x3 frames for a 10-sec video
save_dir = "visual_att/attention_maps/" # store attention maps
original_dir = "visual_att/original/" # store video frames
for num in range(len(test_order)):
print(num)
data = dataset[test_order[num]]
x = data.split("&")
# extract video frames
video_index = os.path.join(raw_video_dir, x[1] + '.mp4')
vid = imageio.get_reader(video_index, 'ffmpeg')
vid_len = len(vid)
frame_interval = int(vid_len / t)
frame_num = video_frame_sample(frame_interval, t, sample_num)
imgs = []
for i, im in enumerate(vid):
x_im = cv2.resize(im, (224, 224))
imgs.append(x_im)
vid.close()
cc = 0
for n in frame_num:
extract_frames[cc, :, :, :] = imgs[n]
cc += 1
# process generated attention maps
att = att_weight[num, :, :, :]
att = normlize(att, 0, 255)
att_scaled = np.zeros((10, 224, 224))
for k in range(att.shape[0]):
att_scaled[k, :, :] = cv2.resize(att[k, :, :], (224, 224)) # scaling attention maps
att_t = np.repeat(att_scaled, 16, axis = 0) # 1-sec segment only has 1 attention map. Here, repeat 16 times to generate 16 maps for a 1-sec video
heat_maps = np.repeat(att_t.reshape(160, 224, 224, 1), 3, axis = -1)
c += 1
att_dir = save_dir + x[1]
ori_dir = original_dir + x[1]
if not os.path.exists(att_dir):
os.makedirs(att_dir)
if not os.path.exists(ori_dir):
os.makedirs(ori_dir)
for idx in range(160):
heat_map = heat_maps[idx, :, :, 0]
im = extract_frames[idx, :, :, :]
im = im[:, :, (2, 1, 0)]
heatmap = cv2.applyColorMap(np.uint8(heat_map), cv2.COLORMAP_JET)
att_frame = heatmap * 0.2 + np.uint8(im) * 0.6
n = "%04d" % idx
vid_index = os.path.join(att_dir, 'pic' + n + '.jpg')
cv2.imwrite(vid_index, att_frame)
ori_frame = np.uint8(im)
ori_index = os.path.join(ori_dir, 'ori' + n + '.jpg')
cv2.imwrite(ori_index, ori_frame)