-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
267 lines (219 loc) · 8.99 KB
/
utils.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
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
import os
import cv2
import numpy as np
import shutil
import scipy.misc
import pdb
import torch
from config import *
def sigmoid(x, derivative=False):
sigm = 1. / (1. + np.exp(-x))
if derivative:
return sigm * (1. - sigm)
return sigm
def normalize_map(s_map):
# normalize the salience map (as done in MIT code)
norm_s_map = (s_map - np.min(s_map))/((np.max(s_map)-np.min(s_map))*1.0)
return norm_s_map
def postprocess_prediction(prediction, size=None, printinfo=False):
"""
Postprocess saliency maps by resizing and applying gaussian blurringself.
args:
prediction: numpy array with saliency postprocess_prediction
size: original (H,W) of the image
returns:
numpy array with saliency map normalized 0-255 (int8)
"""
prediction = prediction - np.min(prediction)
# pdb.set_trace()
# prediction = prediction - np.mean(prediction)
# prediction[prediction<0] = 0
if printinfo:
print('max %.4f min %.4f'%(np.max(prediction), np.min(prediction)))
if np.max(prediction) != 0:
saliency_map = (prediction/np.max(prediction) * 255).astype(np.uint8)
else:
saliency_map = prediction.astype(np.uint8)
if size is None:
size = SALGAN_RESIZE
# resize back to original size
saliency_map = cv2.GaussianBlur(saliency_map, (7, 7), 0)
saliency_map = cv2.resize(saliency_map, (size[1], size[0]), interpolation=cv2.INTER_CUBIC)
# saliency_map = cv2.resize(saliency_map, (size[1], size[0]), interpolation=cv2.INTER_CUBIC)
# saliency_map = cv2.GaussianBlur(saliency_map, (7, 7), 0)
# clip again
# saliency_map = np.clip(saliency_map, 0, 255)
if np.max(saliency_map)!=0:
saliency_map = saliency_map.astype('float') / np.max(saliency_map) * 255.
else:
print('Zero saliency map.')
return saliency_map
def postprocess_prediction_thm(prediction, size=None):
"""
Postprocess saliency maps by resizing and applying gaussian blurringself.
args:
prediction: numpy array with saliency postprocess_prediction
size: original (H,W) of the image
returns:
numpy array with saliency map normalized 0-255 (int8)
"""
# prediction = prediction - np.min(prediction)
prediction = prediction - np.mean(prediction)
prediction[prediction<0] = 0
print('max %.4f min %.4f'%(np.max(prediction), np.min(prediction)))
if np.max(prediction) != 0:
saliency_map = (prediction/np.max(prediction) * 255).astype(np.uint8)
else:
saliency_map = prediction.astype(np.uint8)
if size is None:
size = SALGAN_RESIZE
# resize back to original size
saliency_map = cv2.GaussianBlur(saliency_map, (7, 7), 0)
saliency_map = cv2.resize(saliency_map, (size[1], size[0]), interpolation=cv2.INTER_CUBIC)
# saliency_map = cv2.resize(saliency_map, (size[1], size[0]), interpolation=cv2.INTER_CUBIC)
# saliency_map = cv2.GaussianBlur(saliency_map, (7, 7), 0)
# clip again
# saliency_map = np.clip(saliency_map, 0, 255)
if np.max(saliency_map)!=0:
saliency_map = saliency_map.astype('float') / np.max(saliency_map) * 255.
else:
print('Zero saliency map.')
return saliency_map
def postprocess_prediction_salgan(prediction, size=None):
"""
Postprocess saliency maps by resizing and applying gaussian blurringself.
args:
prediction: numpy array with saliency postprocess_prediction
size: original (H,W) of the image
returns:
numpy array with saliency map normalized 0-255 (int8)
"""
# prediction = prediction - np.min(prediction) # makes no difference
print('max %.4f min %.4f'%(np.max(prediction), np.min(prediction)))
saliency_map = (prediction * 255).astype(np.uint8)
blur_size = 5
# resize back to original size
saliency_map = cv2.resize(saliency_map, (size[1], size[0]), interpolation=cv2.INTER_CUBIC)
# blur
saliency_map = cv2.GaussianBlur(saliency_map, (blur_size, blur_size), 0)
# clip again
saliency_map = np.clip(saliency_map, 0, 255)
return saliency_map
# preserve aspect ratio
def postprocess_hd_prediction(prediction, size=None):
"""
Postprocess saliency maps by resizing and applying gaussian blurringself.
args:
prediction: numpy array with saliency postprocess_prediction
size: original (H,W) of the image
returns:
numpy array with saliency map normalized 0-255 (int8)
"""
print('max %.4f min %.4f'%(np.max(prediction), np.min(prediction)))
saliency_map = (prediction * 255).astype(np.uint8)
if size is None:
size = SALGAN_RESIZE
# resize back to original size
# saliency_map = cv2.GaussianBlur(saliency_map, (7, 7), 0)
# saliency_map = cv2.resize(saliency_map, (size[1], size[0]), interpolation=cv2.INTER_CUBIC)
saliency_map = cv2.resize(saliency_map, (size[1], size[0]), interpolation=cv2.INTER_CUBIC)
# saliency_map = cv2.GaussianBlur(saliency_map, (7, 7), 0) # hd_map has its own gaussian blur process
# clip again
# saliency_map = np.clip(saliency_map, 0, 255)
if np.max(saliency_map)!=0:
saliency_map = saliency_map.astype('float') / np.max(saliency_map) * 255.
else:
print('Zero saliency map.')
return saliency_map
# preserve aspect ratio
def postprocess_prediction_my(pred, shape_r, shape_c):
# pred = sigmoid(pred)
pred = pred.astype('float')
predictions_shape = pred.shape
rows_rate = shape_r*1.0 / predictions_shape[0]
cols_rate = shape_c*1.0 / predictions_shape[1]
pred = pred / np.max(pred) * 255.
if rows_rate > cols_rate:
new_cols = int(predictions_shape[1] * float(shape_r) // predictions_shape[0])
# pred = cv2.resize(pred, (new_cols, shape_r))
pred = scipy.misc.imresize(pred, (shape_r, new_cols))
img = pred[:, ((pred.shape[1] - shape_c) // 2):((pred.shape[1] - shape_c) // 2 + shape_c)]
else:
new_rows = int(predictions_shape[0] * float(shape_c) // predictions_shape[1])
# pred = cv2.resize(pred, (shape_c, new_rows))
pred = scipy.misc.imresize(pred, (new_rows, shape_c))
img = pred[((pred.shape[0] - shape_r) // 2):((pred.shape[0] - shape_r) // 2 + shape_r),:]
img = scipy.ndimage.filters.gaussian_filter(img, sigma=7)
img = img.astype('float') / np.max(img) * 255
return img
# Method to save trained model
def save_model(net, optim, epoch, p_out, eval_loss, name_model=None, results=None, is_best=False, best_name='best.pt'):
if name_model is None:
name_model = epoch
state_dict = net.state_dict()
for key in state_dict.keys():
state_dict[key] = state_dict[key].cpu()
# opt_state_dict = optim.state_dict()
# for key in opt_state_dict.keys():
# opt_state_dict[key] = opt_state_dict[key].cpu()
model_dict = {
'epoch': epoch,
'state_dict': state_dict,
'optimizer': optim.state_dict(),
'eval_loss': eval_loss}
for k in results.keys():
model_dict[k] = results[k].mean
# filepath = os.path.join(p_out,'{}.pt'.format(name_model))
filepath = os.path.join(p_out, name_model)
torch.save(model_dict, filepath)
if is_best:
# shutil.copyfile(filepath, os.path.join(p_out, '{}_best.pt'.format(name_model)))
shutil.copyfile(filepath, os.path.join(p_out, best_name))
# torch.save({
# 'epoch': epoch,
# 'state_dict': state_dict,
# 'optimizer': optim,
# 'eval_loss': eval_loss},
# os.path.join(p_out,'{}_epoch{:02d}.pt'.format(name_model, epoch)))
def get_lr_optimizer( optimizer ):
""" Get learning rate from optimizer."""
for param_group in optimizer.param_groups:
yield param_group['lr']
# -------from dim.py ---------------
def sample_locations(enc, n_samples):
'''Randomly samples locations from localized features.
Used for saving memory.
Args:
enc: Features.
n_samples: Number of samples to draw.
Returns:
torch.Tensor
'''
n_locs = enc.size(2)
batch_size = enc.size(0)
weights = torch.tensor([1. / n_locs] * n_locs, dtype=torch.float)
idx = torch.multinomial(weights, n_samples * batch_size, replacement=True) \
.view(batch_size, n_samples)
enc = enc.transpose(1, 2)
adx = torch.arange(0, batch_size).long()
enc = enc[adx[:, None], idx].transpose(1, 2)
return enc
def sample_locations_my(enc, n_samples):
'''Randomly samples locations from localized features.
Used for saving memory.
Args:
enc: Features. (N, C, H, W)
n_samples: Number of samples to draw.
Returns:
torch.Tensor
'''
enc = enc.view(enc.size(0), enc.size(1), -1)
n_locs = enc.size(2)
batch_size = enc.size(0)
weights = torch.tensor([1. / n_locs] * n_locs, dtype=torch.float)
idx = torch.multinomial(weights, n_samples * batch_size, replacement=True) \
.view(batch_size, n_samples)
enc = enc.transpose(1, 2)
adx = torch.arange(0, batch_size).long()
enc = enc[adx[:, None], idx].transpose(1, 2)
return enc