-
Notifications
You must be signed in to change notification settings - Fork 4
/
cam_2.py
259 lines (210 loc) · 9.61 KB
/
cam_2.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
import tensorflow as tf
tf.python.control_flow_ops = tf
from keras.models import Sequential, model_from_json, load_model
from keras.optimizers import Adam, SGD
from keras.layers.core import Dense, Activation, Flatten, Dropout
from keras.layers.convolutional import Convolution2D
from keras.layers.pooling import MaxPooling2D
from keras.preprocessing.image import *
from keras import __version__ as keras_version
from sklearn.utils import shuffle
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt
import tables
import sys
import cv2
import scipy
import argparse
import h5py
import shutil
import numpy as np
K.set_learning_phase(0) # All operations in test mode
EPSILON = 1e-7
# The order of layers listed in CONV_LAYERS matters in functions below.
CONV_LAYERS = [ 'convolution2d_1', 'convolution2d_2',
'convolution2d_3', 'convolution2d_4', 'convolution2d_5' ]
def normalize_grayscale(image_data):
"""
Normalize the image data with Min-Max scaling to a range of [0.1, 0.9]
:param image_data: The image data to be normalized
:return: Normalized image data
"""
img_max = np.max(image_data)
img_min = np.min(image_data)
a = -0.5
b = 0.5
img_normed = a + (b-a)*(image_data - img_min)/(img_max - img_min)
#print(np.max(img_normed))
#print(np.min(img_normed))
return img_normed
def normalize_color(image_data):
"""
Normalize the image data on per channel basis. """
img_normed_color = np.zeros_like(image_data, dtype=float)
for ch in range(image_data.shape[3]):
tmp = normalize_grayscale(image_data[:,:,:,ch])
img_normed_color[:,:,:,ch] = tmp
#print(np.max(img_normed_color))
#print(np.min(img_normed_color))
return img_normed_color
def normalize(x):
# utility function to normalize a tensor by its L2 norm
return x / (K.sqrt(K.mean(K.square(x))) + 1e-5)
def grad_cam_loss(x, angle):
'''
Return the gradient value, but when angle is very small, we care about those values
contribute to the result of its being small, thus the last "else" of 1/x.
TODO: It seems this method cannot deal with negative values well. Say in small angle
case that postive and negative values average out each other.
'''
# threshold was setting to a value other than 0 degree
#threshold_degree = 3.0
threshold_degree = 0.0
if angle > threshold_degree * scipy.pi / 180.0:
return x
elif angle < -threshold_degree * scipy.pi / 180.0:
return -x
else:
# Avoid div-by-0.
x = x + EPSILON
return (1.0/x) * np.sign(angle)
def visualize_class_activation_map(gradients_function, img_path, output_path):
#print('DEBUG_5: ', img_path)
original_img = cv2.imread(img_path, 1)
width, height, _ = original_img.shape
# !!!IMPORTANT!!!
# Recorded image using OpenCV (BGR), convert to RGB before feeding into network.
# The replayed image has predicted angle a little bit different from recorded value, this is
# probably because img compression/de-compression in recording and replay.
img_rgb = cv2.cvtColor(original_img, cv2.COLOR_BGR2RGB)
img = normalize_color(img_rgb[None,:,:,:])
img = img.reshape([1,66,200,3])
cam_list = []
layer = 0
# Loop through all conv layers, get cam for each
for gf in gradients_function:
layer+=1
interested_layer_outputs, grads_val, angle = gf([img])
#print('DEBUG_6: ', interested_layer_outputs.shape)
#print('DEBUG_7: ', grads_val.shape)
# Sanity check angle vs. directly prediction from model steering_angle. They should match.
#steering_angle = float(model.predict(img, batch_size=1))
#print("predicted angle = ", angle)
#print("steering_angle = ", steering_angle)
#class_weights = np.mean(grads_val, axis=(0,1))
# Evaluate the angle to determine the weights
class_weights = grad_cam_loss(grads_val, angle)
#print('DEBUG_8 class_weights.shape =',class_weights.shape)
#Create the class activation map.
cam = np.zeros(dtype = np.float32, shape = interested_layer_outputs.shape)
# Element-wise muliplication
cam = class_weights*interested_layer_outputs
#print("DEBUG_9 cam.shape = ", cam.shape)
# Average among the number of filters in this layer
cam = np.mean(cam, axis = (2))
#print("DEBUG_10 cam.shape = ", cam.shape)
#Bug? Should use abs(cam) before scaling to colormap
#cam /= np.max(cam)
cam /= np.max(np.abs(cam))
cam_list.append(cam)
cam = cv2.resize(cam, (height, width))
heatmap = cv2.applyColorMap(np.uint8(255*cam), cv2.COLORMAP_JET)
heatmap[np.where(cam < 0.2)] = 0
new_img = heatmap*0.5 + original_img
cv2.putText(new_img,str(angle),(50,50), cv2.FONT_HERSHEY_PLAIN, 1, (255,255,255), 1)
cv2.imwrite(output_path+'_layer'+str(layer)+'.cam.jpg', new_img)
# Calculate the cumulative version of cam
# Following NVidia new paper: https://arxiv.org/pdf/1704.07911.pdf (Explaining How a Deep
# Neural Network Trained with End-to-End Learning Steers a Car)
#
# TODO: the de-convolution part is not implemented, but using a cv2.resize() which should have
# caused a lot of issues -- a lot of final cam image of all layers has nothing left in cam.
scaled_cam = np.ones((cam_list[4].shape[0], cam_list[4].shape[1])) #(width, height)
for cam in reversed(cam_list): # cam_list order matters. Appended from layer 1 to 5
# Scale up to current layer's size
scaled_cam = cv2.resize(scaled_cam, (cam.shape[1], cam.shape[0]))# Interesting, np created array row/col is opposite order of cv2
#print('DEBUG 11 ', scaled_cam.shape)
# Element-wise muliplication
scaled_cam = np.multiply(scaled_cam, cam)
#print('DEBUG 12 ', scaled_cam.shape)
# Normalize
scaled_cam /= np.max(np.abs(scaled_cam))
scaled_cam = cv2.resize(scaled_cam, (height, width))
heatmap = cv2.applyColorMap(np.uint8(255*scaled_cam), cv2.COLORMAP_JET)
heatmap[np.where(scaled_cam < 0.2)] = 0
new_img = heatmap*0.5 + original_img
cv2.putText(new_img,str(angle),(50,50), cv2.FONT_HERSHEY_PLAIN, 1, (255,255,255), 1)
cv2.imwrite(output_path+'_layers'+'.cam.jpg', new_img)
def prepare_grad_func(model, CONV_LAYERS):
'''
Prepare the gradients function from NN output to each interested conv layers.
Return the grad_func as a list.
'''
# Get the final output of the model
pred_angle = K.sum(model.layers[-1].output)
#print('DEBUG_1: ', model.layers[-1].output)
#print('DEBUG_2: pred_angle', pred_angle)
# Build layer dictionary
layer_dict = dict([(layer.name, layer) for layer in model.layers])
# Loop through interested conv layers to back-annotate the activation
gradients_function = []
for CONV_LAYER in CONV_LAYERS:
interested_layer = layer_dict[CONV_LAYER]
# Gradients from output to interested layer, not sure why it is output as a list [blah],
# thus the [0] at the end
grads = normalize(K.gradients(pred_angle, interested_layer.output)[0])
#print('DEBUG_3: ', K.gradients(pred_angle, interested_layer.output))
#print('DEBUG_4: ', K.gradients(pred_angle, interested_layer.output)[0])
# Feed input, grab output (pred_angle), interested layer output and gradients from
# pred_angle back to interested layer.
gradients_function.append(K.function([model.layers[0].input], [interested_layer.output[0], grads[0], pred_angle]))
return gradients_function
# -------------------------------------
# Restore cover of NVidia end-to-end network
# -------------------------------------
import os
import glob
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Gradients Activation Mapping processing.')
parser.add_argument(
'model',
type=str,
help='Path to model h5 file. Model should be on the same path.'
)
parser.add_argument(
'image_folder',
type=str,
nargs='?',
default='',
help='Path to image folder. This is where the images from the run will be saved.'
)
args = parser.parse_args()
# check that model Keras version is same as local Keras version
f = h5py.File(args.model, mode='r')
model_version = f.attrs.get('keras_version')
keras_version = str(keras_version).encode('utf8')
if model_version != keras_version:
print('You are using Keras version ', keras_version,
', but the model was built using ', model_version)
model = load_model(args.model)
model.summary()
###------ Key section of assistance functions definition ------
gradients_function = prepare_grad_func(model, CONV_LAYERS);
if args.image_folder != '':
print("Work with image folder at {}".format(args.image_folder))
if not os.path.exists(args.image_folder):
print("Image folder doesn't exist!")
else:
in_folder = args.image_folder
out_folder = in_folder + '_cam'
print("Creating image folder at {}".format(out_folder))
if os.path.exists(out_folder):
shutil.rmtree(out_folder)
os.makedirs(out_folder)
for img_file in glob.glob(in_folder+"/*.png"):
print(img_file)
in_folder_str_len = len(in_folder)
trim_img_file = img_file[in_folder_str_len:-4]
visualize_class_activation_map(gradients_function, img_file, out_folder+'/'+trim_img_file)
else:
print("Where are the images stored? Please provide image folder.")