-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathTCN.py
270 lines (225 loc) · 10.5 KB
/
TCN.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
268
269
270
import tensorflow as tf
from PIL import Image
import numpy as np
import imageio
import os
import matplotlib.pyplot as plt
from tensorflow.keras.applications import InceptionV3
from tensorflow.keras.applications.inception_v3 import preprocess_input
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Dense,Conv2D,Flatten,BatchNormalization
import matplotlib.pyplot as plt
alpha=10.0
inception=InceptionV3(include_top=False,pooling="avg")
learning_rate=3e-4
margin=2.0
class CoordinateUtils(object):
@staticmethod
def get_image_coordinates(h, w, normalise):
x_range = tf.range(w, dtype=tf.float32)
y_range = tf.range(h, dtype=tf.float32)
if normalise:
x_range = (x_range / (w - 1)) * 2 - 1
y_range = (y_range / (h - 1)) * 2 - 1
image_x = tf.repeat(tf.expand_dims(x_range,0),h, 0)
image_y = tf.transpose(tf.repeat(tf.expand_dims(y_range,0),w, 0))
return image_x, image_y
class spatial_softmax(tf.keras.Model):
def __init__(self, temperature=None, normalise=True):
super(spatial_softmax,self).__init__()
self.temperature = tf.ones(1) if temperature is None else tf.tensor([temperature])
self.normalise = normalise
def call(self,inputs):
N,H,W,C = inputs.shape
features = tf.reshape(tf.transpose(inputs, [0, 3, 1, 2]), [N * C, H * W])
softmax = tf.nn.softmax(features)
softmax = tf.reshape(softmax, [N, C, H, W])
softmax = tf.expand_dims(softmax, -1)
# Convert image coords to shape [H, W, 1, 2]
image_x, image_y = CoordinateUtils.get_image_coordinates(H, W, normalise=self.normalise)
image_coords = tf.concat([tf.expand_dims(image_x,-1), tf.expand_dims(image_y,-1)],-1)
# Multiply (with broadcasting) and reduce over image dimensions to get the result
# of shape [N, C, 2]
spatial_soft_argmax = tf.reduce_sum(softmax * image_coords,[2, 3])
return spatial_soft_argmax
def normalize(output_embedding):
buffer = tf.math.pow(output_embedding, 2)
normp = tf.math.reduce_sum(buffer, 1,keepdims=True) + 1e-10
normalization_constant = tf.math.sqrt(normp)
output = tf.math.divide(output_embedding, normalization_constant)
return output
def get_embeddings(input_frame):
input_frame=preprocess_input(input_frame)
output_embedding=inception.predict(input_frame)
return output_embedding
class TCNModel(tf.keras.Model):
def __init__(self,temperature=None, normalise=True):
super(TCNModel,self).__init__()
self.layer_1 = Conv2D(100,kernel_size=3,strides=1,activation='relu')
self.layer_2 = BatchNormalization()
self.layer_3 = Conv2D(20,kernel_size=3,strides=1,activation='relu')
self.layer_4 = BatchNormalization()
self.layer_5 = spatial_softmax(temperature=None, normalise=True)
self.layer_6 = Flatten()
self.layer_7 = Dense(32, activation='relu')
def call(self,inputs):
x = self.layer_3(self.layer_2(self.layer_1(inputs)))
x = self.layer_6(self.layer_5(self.layer_4(x)))
x = self.layer_7(x)
x = normalize(x) * alpha
return x
def ls(path):
# returns list of files in directory without hidden ones.
return [p for p in os.listdir(path) if p[0] != '.']
def _resize_frame(frame, out_size):
image = Image.fromarray(frame)
image = image.resize((299,299))
scaled = np.array(image, dtype=np.float32) / 255
return scaled
def read_video(filepath, frame_size):
imageio_video = imageio.read(filepath)
snap_length = len(imageio_video)
frames = np.zeros((snap_length,*frame_size))
for i, frame in enumerate(imageio_video):
frame=_resize_frame(frame, frame_size)
frames[i, :, :, :] = frame
return frames
class SingleViewTripletBuilder(object):
def __init__(self, video_directory, image_size):
self.frame_size = image_size
self._read_video_dir(video_directory)
self._count_frames()
# The negative example has to be from outside the buffer window. Taken from both sides of the frame.
self.positive_frame_margin = 10
self.negative_frame_margin = 30
self.video_index = 0
def _read_video_dir(self, video_directory):
self._video_directory = video_directory
self._anchor_directory=os.path.join(self._video_directory, 'Anchor')
self._positive_directory=os.path.join(self._video_directory, 'Positive')
anchor_filenames = ls(self._anchor_directory)
positive_filenames = ls(self._positive_directory)
self.anchor_video_paths = [os.path.join(self._anchor_directory, f) for f in anchor_filenames]
self.positive_video_paths = [os.path.join(self._positive_directory, f) for f in positive_filenames]
self.video_count = len(self.anchor_video_paths)
def _count_frames(self):
frame_lengths = np.array([len(imageio.read(p)) for p in self.anchor_video_paths])
self.frame_lengths = frame_lengths
def get_video(self, index):
print(self.anchor_video_paths[index],self.positive_video_paths[index])
return (read_video(self.anchor_video_paths[index], self.frame_size),read_video(self.positive_video_paths[index], self.frame_size))
def sample_triplet(self, anchor_video,positive_video):
anchor_index = self.sample_anchor_frame_index()
positive_index = anchor_index
negative_index = self.sample_negative_frame_index(anchor_index)
anchor_frame = anchor_video[anchor_index]
positive_frame = positive_video[positive_index]
negative_frame = anchor_video[negative_index]
return (anchor_frame,positive_frame,negative_frame)
def build_set(self,batch_size):
triplets = []
triplets = np.zeros((batch_size,3,299,299,3))
for i in range(0, batch_size):
anchor_video,positive_video = self.get_video(self.video_index)
anchor_frame, positive_frame, negative_frame = self.sample_triplet(anchor_video,positive_video)
triplets[i, 0, :, :, :] = anchor_frame
triplets[i, 1, :, :, :] = positive_frame
triplets[i, 2, :, :, :] = negative_frame
fig = plt.figure(figsize=(12,12))
ax1 = fig.add_subplot(2,2,1)
ax1.imshow(anchor_frame)
ax2 = fig.add_subplot(2,2,2)
ax2.imshow(positive_frame)
ax3 = fig.add_subplot(2,2,3)
ax3.imshow(negative_frame)
plt.show()
self.video_index = (self.video_index + 1) % self.video_count
# Second argument is labels. Not used.
return triplets
def sample_anchor_frame_index(self):
arange = np.arange(0, self.frame_lengths[self.video_index])
return np.random.choice(arange)
def negative_frame_indices(self, anchor_index):
video_length = self.frame_lengths[self.video_index]
lower_bound = 0
upper_bound = max(0, anchor_index - self.negative_frame_margin)
range1 = np.arange(lower_bound, upper_bound)
lower_bound = min(anchor_index + self.negative_frame_margin, video_length)
upper_bound = video_length
range2 = np.arange(lower_bound, upper_bound)
return np.concatenate([range1, range2])
def sample_negative_frame_index(self, anchor_index):
return np.random.choice(self.negative_frame_indices(anchor_index))
def distance(x1, x2):
diff = ((abs(x1 - x2)**2).sum(dim=1))
return diff
def validate(tcn):
# Run model on validation data and log results
tcn=TCNModel()
IMAGE_SIZE = (299, 299,3)
triplet_builder = SingleViewTripletBuilder("Validation", IMAGE_SIZE)
batch_size=64
validation_dataset = triplet_builder.build_set(batch_size)
correct_with_margin = 0
correct_without_margin = 0
anchor_frames = validation_dataset[:, 0, :, :, :]
positive_frames = validation_dataset[:, 1, :, :, :]
negative_frames = validation_dataset[:, 2, :, :, :]
anchor_embeddings = get_embeddings(anchor_frames)
positive_embeddings = get_embeddings(positive_frames)
negative_embeddings = get_embeddings(negative_frames)
anchor_output = tcn(anchor_embeddings)
positive_output = tcn(positive_embeddings)
negative_output = tcn(negative_embeddings)
d_positive = distance(anchor_output, positive_output)
d_negative = distance(anchor_output, negative_output)
for i in range(64):
print('D_positive',d_positive[i],'D_negative',d_negative[i])
fig = plt.figure(figsize=(12,12))
ax1 = fig.add_subplot(2,2,1)
ax1.imshow(anchor_frames[i])
ax2 = fig.add_subplot(2,2,2)
ax2.imshow(positive_frames[i])
ax3 = fig.add_subplot(2,2,3)
ax3.imshow(negative_frames[i])
plt.show()
correct_with_margin +=1*((d_positive + margin) < d_negative)
correct_without_margin += 1*(d_positive < d_negative)
print("Validation score correct with margin {with_margin}/{total} and without margin {without_margin}{total}".format(with_margin=tf.reduce_sum(correct_with_margin),without_margin=tf.reduce_sum(correct_without_margin),total=len(validation_dataset)))
def train_step(anchor_embeddings,positive_embeddings,negative_embeddings,tcn):
optimizer_policy = tf.keras.optimizers.Adam(learning_rate=learning_rate)
with tf.GradientTape() as tape:
anchor_output = tcn(anchor_embeddings)
positive_output = tcn(positive_embeddings)
negative_output = tcn(negative_embeddings)
d_positive = distance(anchor_output, positive_output)
d_negative = distance(anchor_output, negative_output)
loss=margin + d_positive - d_negative
grads=tape.gradient(loss,tcn.trainable_variables)
optimizer_Q_function.apply_gradients(zip(grads, tcn.trainable_variables))
return loss
def train_tcn():
tcn=TCNModel()
epochs=1000
IMAGE_SIZE = (299, 299,3)
triplet_builder = SingleViewTripletBuilder("Training", IMAGE_SIZE)
ITERATE_OVER_TRIPLETS = 5
batch_size=128
for epoch in range(epochs):
training_dataset = triplet_builder.build_set(batch_size)
anchor_frames = training_dataset[:, 0, :, :, :]
positive_frames = training_dataset[:, 1, :, :, :]
negative_frames = training_dataset[:, 2, :, :, :]
anchor_embeddings = get_embeddings(anchor_frames)
positive_embeddings = get_embeddings(positive_frames)
negative_embeddings = get_embeddings(negative_frames)
loss = 0
if epoch%100==0 and epoch!=0 and learning_rate>3e-5:
learning_rate=learning_rate*0.1
for _ in range(0, ITERATE_OVER_TRIPLETS):
loss+=train_step(anchor_embeddings,positive_embeddings,negative_embeddings,tcn)
print("Epoch:{} Loss:{}".format(epoch,np.mean(loss)))
tcn.save('tcn_{}.h5'.format(epoch))
if epoch % 10 == 0 and epoch!=0:
validate(tcn)
train_tcn()