frame_level_models.py
20.5 KB
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
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
# Copyright 2016 Google Inc. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS-IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Contains a collection of models which operate on variable-length sequences."""
import math
import model_utils as utils
import models
import tensorflow as tf
from tensorflow import flags
import tensorflow.contrib.slim as slim
import video_level_models
FLAGS = flags.FLAGS
flags.DEFINE_integer("iterations", 30, "Number of frames per batch for DBoF.")
flags.DEFINE_bool("dbof_add_batch_norm", True,
"Adds batch normalization to the DBoF model.")
flags.DEFINE_bool(
"sample_random_frames", True,
"If true samples random frames (for frame level models). If false, a random"
"sequence of frames is sampled instead.")
flags.DEFINE_integer("dbof_cluster_size", 8192,
"Number of units in the DBoF cluster layer.")
flags.DEFINE_integer("dbof_hidden_size", 1024,
"Number of units in the DBoF hidden layer.")
flags.DEFINE_string(
"dbof_pooling_method", "max",
"The pooling method used in the DBoF cluster layer. "
"Choices are 'average' and 'max'.")
flags.DEFINE_string(
"dbof_activation", "sigmoid",
"The nonlinear activation method for cluster and hidden dense layer, e.g., "
"sigmoid, relu6, etc.")
flags.DEFINE_string(
"video_level_classifier_model", "MoeModel",
"Some Frame-Level models can be decomposed into a "
"generalized pooling operation followed by a "
"classifier layer")
flags.DEFINE_integer("lstm_cells", 512, "Number of LSTM cells.")
flags.DEFINE_integer("lstm_layers", 2, "Number of LSTM layers.")
class DbofModel(models.BaseModel):
"""Creates a Deep Bag of Frames model.
The model projects the features for each frame into a higher dimensional
'clustering' space, pools across frames in that space, and then
uses a configurable video-level model to classify the now aggregated features.
The model will randomly sample either frames or sequences of frames during
training to speed up convergence.
"""
ACT_FN_MAP = {
"sigmoid": tf.nn.sigmoid,
"relu6": tf.nn.relu6,
}
def create_model(self,
model_input,
vocab_size,
num_frames,
iterations=None,
add_batch_norm=None,
sample_random_frames=None,
cluster_size=None,
hidden_size=None,
is_training=True,
**unused_params):
"""See base class.
Args:
model_input: A 'batch_size' x 'max_frames' x 'num_features' matrix of
input features.
vocab_size: The number of classes in the dataset.
num_frames: A vector of length 'batch' which indicates the number of
frames for each video (before padding).
iterations: the number of frames to be sampled.
add_batch_norm: whether to add batch norm during training.
sample_random_frames: whether to sample random frames or random sequences.
cluster_size: the output neuron number of the cluster layer.
hidden_size: the output neuron number of the hidden layer.
is_training: whether to build the graph in training mode.
Returns:
A dictionary with a tensor containing the probability predictions of the
model in the 'predictions' key. The dimensions of the tensor are
'batch_size' x 'num_classes'.
"""
iterations = iterations or FLAGS.iterations
add_batch_norm = add_batch_norm or FLAGS.dbof_add_batch_norm
random_frames = sample_random_frames or FLAGS.sample_random_frames
cluster_size = cluster_size or FLAGS.dbof_cluster_size
hidden1_size = hidden_size or FLAGS.dbof_hidden_size
act_fn = self.ACT_FN_MAP.get(FLAGS.dbof_activation)
assert act_fn is not None, ("dbof_activation is not valid: %s." %
FLAGS.dbof_activation)
num_frames = tf.cast(tf.expand_dims(num_frames, 1), tf.float32)
if random_frames:
model_input = utils.SampleRandomFrames(model_input, num_frames,
iterations)
else:
model_input = utils.SampleRandomSequence(model_input, num_frames,
iterations)
max_frames = model_input.get_shape().as_list()[1]
feature_size = model_input.get_shape().as_list()[2]
reshaped_input = tf.reshape(model_input, [-1, feature_size])
tf.compat.v1.summary.histogram("input_hist", reshaped_input)
if add_batch_norm:
reshaped_input = slim.batch_norm(reshaped_input,
center=True,
scale=True,
is_training=is_training,
scope="input_bn")
cluster_weights = tf.compat.v1.get_variable(
"cluster_weights", [feature_size, cluster_size],
initializer=tf.random_normal_initializer(stddev=1 /
math.sqrt(feature_size)))
tf.compat.v1.summary.histogram("cluster_weights", cluster_weights)
activation = tf.matmul(reshaped_input, cluster_weights)
if add_batch_norm:
activation = slim.batch_norm(activation,
center=True,
scale=True,
is_training=is_training,
scope="cluster_bn")
else:
cluster_biases = tf.compat.v1.get_variable(
"cluster_biases", [cluster_size],
initializer=tf.random_normal_initializer(stddev=1 /
math.sqrt(feature_size)))
tf.compat.v1.summary.histogram("cluster_biases", cluster_biases)
activation += cluster_biases
activation = act_fn(activation)
tf.compat.v1.summary.histogram("cluster_output", activation)
activation = tf.reshape(activation, [-1, max_frames, cluster_size])
activation = utils.FramePooling(activation, FLAGS.dbof_pooling_method)
hidden1_weights = tf.compat.v1.get_variable(
"hidden1_weights", [cluster_size, hidden1_size],
initializer=tf.random_normal_initializer(stddev=1 /
math.sqrt(cluster_size)))
tf.compat.v1.summary.histogram("hidden1_weights", hidden1_weights)
activation = tf.matmul(activation, hidden1_weights)
if add_batch_norm:
activation = slim.batch_norm(activation,
center=True,
scale=True,
is_training=is_training,
scope="hidden1_bn")
else:
hidden1_biases = tf.compat.v1.get_variable(
"hidden1_biases", [hidden1_size],
initializer=tf.random_normal_initializer(stddev=0.01))
tf.compat.v1.summary.histogram("hidden1_biases", hidden1_biases)
activation += hidden1_biases
activation = act_fn(activation)
tf.compat.v1.summary.histogram("hidden1_output", activation)
aggregated_model = getattr(video_level_models,
FLAGS.video_level_classifier_model)
return aggregated_model().create_model(model_input=activation,
vocab_size=vocab_size,
**unused_params)
class LstmModel(models.BaseModel):
"""Creates a model which uses a stack of LSTMs to represent the video."""
def create_model(self, model_input, vocab_size, num_frames, **unused_params):
"""See base class.
Args:
model_input: A 'batch_size' x 'max_frames' x 'num_features' matrix of
input features.
vocab_size: The number of classes in the dataset.
num_frames: A vector of length 'batch' which indicates the number of
frames for each video (before padding).
Returns:
A dictionary with a tensor containing the probability predictions of the
model in the 'predictions' key. The dimensions of the tensor are
'batch_size' x 'num_classes'.
"""
lstm_size = FLAGS.lstm_cells
number_of_layers = FLAGS.lstm_layers
stacked_lstm = tf.contrib.rnn.MultiRNNCell([
tf.contrib.rnn.BasicLSTMCell(lstm_size, forget_bias=1.0)
for _ in range(number_of_layers)
])
_, state = tf.nn.dynamic_rnn(stacked_lstm,
model_input,
sequence_length=num_frames,
dtype=tf.float32)
aggregated_model = getattr(video_level_models,
FLAGS.video_level_classifier_model)
return aggregated_model().create_model(model_input=state[-1].h,
vocab_size=vocab_size,
**unused_params)
class GruModel(models.BaseModel):
def create_model(self, model_input, vocab_size, num_frames, is_training=True, **unused_params):
"""Creates a model which uses a stack of GRUs to represent the video.
Args:
model_input: A 'batch_size' x 'max_frames' x 'num_features' matrix of
input features.
vocab_size: The number of classes in the dataset.
num_frames: A vector of length 'batch' which indicates the number of
frames for each video (before padding).
Returns:
A dictionary with a tensor containing the probability predictions of the
model in the 'predictions' key. The dimensions of the tensor are
'batch_size' x 'num_classes'.
"""
gru_size = FLAGS.lstm_cells
number_of_layers = FLAGS.lstm_layers
backward = False
random_frames = False
iterations = 30
if random_frames:
num_frames_2 = tf.cast(tf.expand_dims(num_frames, 1), tf.float32)
model_input = utils.SampleRandomFrames(model_input, num_frames_2,
iterations)
if backward:
model_input = tf.reverse_sequence(model_input, num_frames, seq_axis=1)
stacked_GRU = tf.contrib.rnn.MultiRNNCell(
[
tf.contrib.rnn.GRUCell(gru_size)
for _ in range(number_of_layers)
], state_is_tuple=False)
loss = 0.0
with tf.variable_scope("RNN"):
outputs, state = tf.nn.dynamic_rnn(stacked_GRU, model_input,
sequence_length=num_frames,
dtype=tf.float32)
aggregated_model = getattr(video_level_models,
'MoeModel')
return aggregated_model().create_model(
model_input=state,
vocab_size=vocab_size,
is_training=is_training,
**unused_params)
class FrameLevelLogisticModel(models.BaseModel):
"""Creates a logistic classifier over the aggregated frame-level features."""
def create_model(self, model_input, vocab_size, num_frames, **unused_params):
"""See base class.
This class is intended to be an example for implementors of frame level
models. If you want to train a model over averaged features it is more
efficient to average them beforehand rather than on the fly.
Args:
model_input: A 'batch_size' x 'max_frames' x 'num_features' matrix of
input features.
vocab_size: The number of classes in the dataset.
num_frames: A vector of length 'batch' which indicates the number of
frames for each video (before padding).
Returns:
A dictionary with a tensor containing the probability predictions of the
model in the 'predictions' key. The dimensions of the tensor are
'batch_size' x 'num_classes'.
"""
num_frames = tf.cast(tf.expand_dims(num_frames, 1), tf.float32)
feature_size = model_input.get_shape().as_list()[2]
denominators = tf.reshape(tf.tile(num_frames, [1, feature_size]),
[-1, feature_size])
avg_pooled = tf.reduce_sum(model_input, axis=[1]) / denominators
output = slim.fully_connected(avg_pooled,
vocab_size,
activation_fn=tf.nn.sigmoid,
weights_regularizer=slim.l2_regularizer(1e-8))
return {"predictions": output}
class NetVLAD_NonLocal_types():
def __init__(self, feature_size,max_frames,cluster_size, add_batch_norm, is_training):
self.feature_size = feature_size
self.max_frames = max_frames
self.is_training = is_training
self.add_batch_norm = add_batch_norm
self.cluster_size = cluster_size
def forward(self,reshaped_input):
cluster_weights = tf.get_variable("cluster_weights",
[self.feature_size, self.cluster_size],
initializer = tf.random_normal_initializer(stddev=1 / math.sqrt(self.feature_size)))
tf.summary.histogram("cluster_weights", cluster_weights)
activation = tf.matmul(reshaped_input, cluster_weights)
if self.add_batch_norm:
activation = slim.batch_norm(activation, center=True, scale=True, is_training=self.is_training, scope="cluster_bn")
else:
cluster_biases = tf.get_variable("cluster_biases",
[cluster_size],
initializer = tf.random_normal_initializer(stddev=1 / math.sqrt(self.feature_size)))
tf.summary.histogram("cluster_biases", cluster_biases)
activation += cluster_biases
activation = tf.nn.softmax(activation)
tf.summary.histogram("cluster_output", activation)
activation = tf.reshape(activation, [-1, self.max_frames, self.cluster_size])
a_sum = tf.reduce_sum(activation,-2,keep_dims=True)
cluster_weights2 = tf.get_variable("cluster_weights2",
[1,self.feature_size, self.cluster_size],
initializer = tf.random_normal_initializer(stddev=1 / math.sqrt(self.feature_size)))
a = tf.multiply(a_sum,cluster_weights2)
activation = tf.transpose(activation,perm=[0,2,1])
reshaped_input = tf.reshape(reshaped_input,[-1,self.max_frames,self.feature_size])
vlad = tf.matmul(activation,reshaped_input)
vlad = tf.transpose(vlad,perm=[0,2,1])
vlad = tf.subtract(vlad,a)
vlad = tf.transpose(vlad,perm=[0,2,1])
vlad = tf.reshape(vlad, [-1, self.feature_size])
vlad_softmax = self.embedgaussian_relation(vlad, 1/float(64))
nonlocal_g = tf.get_variable("nonlocal_g",
[self.feature_size, self.cluster_size],
initializer = tf.random_normal_initializer(stddev=1 / math.sqrt(self.feature_size)))
nonlocal_out = tf.get_variable("nonlocal_out",
[self.cluster_size, self.feature_size],
initializer = tf.random_normal_initializer(stddev=1 / math.sqrt(self.cluster_size)))
vlad_g = tf.matmul(vlad, nonlocal_g)
vlad_g = tf.reshape(vlad_g, [-1, self.cluster_size, self.cluster_size])
vlad_g = tf.matmul(vlad_softmax, vlad_g)
vlad_g = tf.reshape(vlad_g, [-1, self.cluster_size])
vlad_g = tf.matmul(vlad_g, nonlocal_out)
vlad_g = tf.reshape(vlad_g, [-1, self.cluster_size, self.feature_size])
vlad = tf.reshape(vlad, [-1, self.cluster_size, self.feature_size])
vlad = vlad + vlad_g
vlad = tf.transpose(vlad,perm=[0,2,1])
vlad = tf.nn.l2_normalize(vlad,1) # [b,f,c]
vlad = tf.reshape(vlad,[-1,self.cluster_size*self.feature_size])
vlad = tf.nn.l2_normalize(vlad,1)
return vlad
def embedgaussian_relation(self, input_, temp=1/float(32)):
nonlocal_theta = tf.get_variable("nonlocal_theta",
[self.feature_size, self.cluster_size],
initializer = tf.random_normal_initializer(stddev=1 / math.sqrt(self.feature_size)))
nonlocal_phi = tf.get_variable("nonlocal_phi",
[self.feature_size, self.cluster_size],
initializer = tf.random_normal_initializer(stddev=1 / math.sqrt(self.feature_size)))
vlad_theta = tf.matmul(input_, nonlocal_theta)
vlad_phi = tf.matmul(input_, nonlocal_phi)
vlad_theta = tf.reshape(vlad_theta, [-1, self.cluster_size, self.cluster_size])
vlad_phi = tf.reshape(vlad_phi, [-1, self.cluster_size, self.cluster_size])
vlad_softmax = tf.nn.softmax(temp * tf.matmul(vlad_theta, tf.transpose(vlad_phi,perm=[0,2,1])))
return vlad_softmax
class NetVLADModelLF(models.BaseModel):
"""Creates a NetVLAD based model.
Args:
model_input: A 'batch_size' x 'max_frames' x 'num_features' matrix of
input features.
vocab_size: The number of classes in the dataset.
num_frames: A vector of length 'batch' which indicates the number of
frames for each video (before padding).
Returns:
A dictionary with a tensor containing the probability predictions of the
model in the 'predictions' key. The dimensions of the tensor are
'batch_size' x 'num_classes'.
"""
def create_model(self,model_input,vocab_size,num_frames,iterations=None,add_batch_norm=None,sample_random_frames=None,cluster_size=None,hidden_size=None,is_training=True,**unused_params):
iterations = 300
add_batch_norm = True
random_frames = True
cluster_size = 64
hidden1_size = 1024
relu = True
dimred = -1
gating = True
remove_diag = False
num_frames = tf.cast(tf.expand_dims(num_frames, 1), tf.float32)
if random_frames:
model_input = utils.SampleRandomFrames(model_input, num_frames,
iterations)
else:
model_input = utils.SampleRandomSequence(model_input, num_frames,
iterations)
max_frames = model_input.get_shape().as_list()[1]
feature_size = model_input.get_shape().as_list()[2]
reshaped_input = tf.reshape(model_input, [-1, feature_size])
video_NetVLAD = NetVLAD_NonLocal_types(1024,int(max_frames),int(cluster_size), add_batch_norm, is_training)
audio_NetVLAD = NetVLAD_NonLocal_types(128,int(max_frames),int(cluster_size/2), add_batch_norm, is_training)
if add_batch_norm:# and not lightvlad:
reshaped_input = slim.batch_norm(
reshaped_input,
center=True,
scale=True,
is_training=is_training,
scope="input_bn")
with tf.variable_scope("video_VLAD"):
vlad_video = video_NetVLAD.forward(reshaped_input[:,0:1024])
with tf.variable_scope("audio_VLAD"):
vlad_audio = audio_NetVLAD.forward(reshaped_input[:,1024:])
vlad = tf.concat([vlad_video, vlad_audio],1)
vlad_dim = vlad.get_shape().as_list()[1]
hidden1_weights = tf.get_variable("hidden1_weights",
[vlad_dim, hidden1_size],
initializer=tf.random_normal_initializer(stddev=1 / math.sqrt(cluster_size)))
activation = tf.matmul(vlad, hidden1_weights)
if add_batch_norm and relu:
activation = slim.batch_norm(
activation,
center=True,
scale=True,
is_training=is_training,
scope="hidden1_bn")
else:
hidden1_biases = tf.get_variable("hidden1_biases",
[hidden1_size],
initializer = tf.random_normal_initializer(stddev=0.01))
tf.summary.histogram("hidden1_biases", hidden1_biases)
activation += hidden1_biases
if relu:
activation = tf.nn.relu6(activation)
if gating:
gating_weights = tf.get_variable("gating_weights_2",
[hidden1_size, hidden1_size],
initializer = tf.random_normal_initializer(stddev=1 / math.sqrt(hidden1_size)))
gates = tf.matmul(activation, gating_weights)
if remove_diag:
#removes diagonals coefficients
diagonals = tf.matrix_diag_part(gating_weights)
gates = gates - tf.multiply(diagonals,activation)
if add_batch_norm:
gates = slim.batch_norm(
gates,
center=True,
scale=True,
is_training=is_training,
scope="gating_bn")
else:
gating_biases = tf.get_variable("gating_biases",
[cluster_size],
initializer = tf.random_normal(stddev=1 / math.sqrt(feature_size)))
gates += gating_biases
gates = tf.sigmoid(gates)
activation = tf.multiply(activation,gates)
aggregated_model = getattr(video_level_models,
'willow_MoeModel')
return aggregated_model().create_model(
model_input=activation,
vocab_size=vocab_size,
is_training=is_training,
**unused_params)