def modelOpt(self):
bn_params = {'decay': 0.999, 'center': True, 'scale': True, 'epsilon': 0.001, 'updates_collections': None,
'is_training': self.is_training}
self.global_step = tf.Variable(0, trainable=False)
self.learning_rate = layers.decayLearningRate(LEARNING_RATE, self.global_step, DECAY_STEPS, DECAY_RATE)
self.towerLogits = tf.placeholder(dtype=tf.float32, shape=[None, self.dataset.frames, 64])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
net_shape = self.towerLogits.get_shape()
net = tf.reshape(self.towerLogits, [BATCH_SIZE, int(net_shape[1]) * int(net_shape[2])])
layer_num = 1
for fully_connected_num in [64]:
net = layers.fc(net, fully_connected_num, name='temporal_FC{}'.format(layer_num),
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE),
normalizer_fn=layers.batchNormalization, normalizer_params=bn_params)
layer_num += 1
self.logits = layers.fc(net, self.dataset.num_classes, activation_fn=None, name='logits')
self.preds = layers.softmax(self.logits)
cross_entropy_loss = layers.reduce_mean(layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
regularization_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
self.loss = cross_entropy_loss + REGULARIZER_SCALE * tf.reduce_sum(regularization_loss)
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss, global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(self.Yoh, self.preds, self.loss,
self.learning_rate)
def createModel(self):
print("CREATING MODEL")
self.global_step = tf.Variable(0, trainable=False)
self.is_training = tf.placeholder_with_default(True, [], name='is_training')
self.learning_rate = layers.decayLearningRate(LEARNING_RATE, self.global_step, DECAY_STEPS, DECAY_RATE)
self.X = tf.placeholder(dtype=tf.float32, shape=[None, self.dataset.frames, self.dataset.h, self.dataset.w, self.dataset.c])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
bn_params = {'decay': 0.999, 'center': True, 'scale': True, 'epsilon': 0.001, 'updates_collections': None, 'is_training': self.is_training}
concated = None
reuse = None
for sequence_image in range(self.dataset.frames):
net = self.vgg(self.X[:, sequence_image], reuse)
net_shape = net.get_shape()
net = tf.reshape(net, [BATCH_SIZE, int(net_shape[1]) * int(net_shape[2]) * int(net_shape[3])])
net = layers.fc(net, 64, name='spatial_FC', reuse=reuse,
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE),
normalizer_fn=layers.batchNormalization, normalizer_params=bn_params)
if concated is None:
concated = tf.expand_dims(net, axis=1)
else:
concated = tf.concat([concated, tf.expand_dims(net, axis=1)], axis=1)
reuse = True
net = concated
net_shape = net.get_shape()
net = tf.reshape(net, [BATCH_SIZE, int(net_shape[1]) * int(net_shape[2])])
layer_num = 1
for fully_connected_num in [64]:
net = layers.fc(net, fully_connected_num, name='temporal_FC{}'.format(layer_num),
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE),
normalizer_fn=layers.batchNormalization, normalizer_params=bn_params)
layer_num += 1
self.logits = layers.fc(net, self.dataset.num_classes, activation_fn=None, name='logits')
self.preds = layers.softmax(self.logits)
cross_entropy_loss = layers.reduce_mean(layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
regularization_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
self.loss = cross_entropy_loss + REGULARIZER_SCALE * tf.reduce_sum(regularization_loss)
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss, global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(self.Yoh, self.preds, self.loss,
self.learning_rate)
def modelOpt(self):
bn_params = {
'decay': 0.999,
'center': True,
'scale': True,
'epsilon': 0.001,
'updates_collections': None,
'is_training': self.is_training
}
self.global_step = tf.Variable(0, trainable=False)
self.learning_rate = layers.decayLearningRate(LEARNING_RATE,
self.global_step,
DECAY_STEPS, DECAY_RATE)
self.towerLogits = tf.placeholder(
dtype=tf.float32, shape=[None, self.dataset.frames, 64])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
net_shape = self.towerLogits.get_shape()
net = tf.reshape(
self.towerLogits,
[BATCH_SIZE, int(net_shape[1]) * int(net_shape[2])])
layer_num = 1
for fully_connected_num in [64]:
net = layers.fc(
net,
fully_connected_num,
name='temporal_FC{}'.format(layer_num),
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE),
normalizer_fn=layers.batchNormalization,
normalizer_params=bn_params)
layer_num += 1
self.logits = layers.fc(net,
self.dataset.num_classes,
activation_fn=None,
name='logits')
self.preds = layers.softmax(self.logits)
cross_entropy_loss = layers.reduce_mean(
layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
regularization_loss = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
self.loss = cross_entropy_loss + REGULARIZER_SCALE * tf.reduce_sum(
regularization_loss)
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss,
global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(
self.Yoh, self.preds, self.loss, self.learning_rate)
def createModelLowMemory(self):
print("CREATING MODEL")
self.global_step = tf.Variable(0, trainable=False)
self.is_training = tf.placeholder_with_default(True, [], name='is_training')
self.learning_rate = layers.decayLearningRate(LEARNING_RATE, self.global_step, DECAY_STEPS, DECAY_RATE)
self.X = tf.placeholder(dtype=tf.float32, shape=[None, self.dataset.frames, self.dataset.h, self.dataset.w, self.dataset.c])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
concated = None
reuse = None
for sequence_image in range(self.dataset.frames):
net = self.vggLowMemory(self.X[:, sequence_image], reuse)
net_shape = net.get_shape()
net = tf.reshape(net, [BATCH_SIZE, int(net_shape[1]) * int(net_shape[2]) * int(net_shape[3])])
net = layers.fc(net, 64, name='spatial_FC', reuse=reuse)
if concated is None:
concated = tf.expand_dims(net, axis=1)
else:
concated = tf.concat([concated, tf.expand_dims(net, axis=1)], axis=1)
reuse = True
net = concated
net_shape = net.get_shape()
net = tf.reshape(net, [BATCH_SIZE, int(net_shape[1]) * int(net_shape[2])])
layer_num = 1
for fully_connected_num in [64]:
net = layers.fc(net, fully_connected_num, name='temporal_FC{}'.format(layer_num))
layer_num += 1
self.logits = layers.fc(net, self.dataset.num_classes, activation_fn=None, name='logits')
self.preds = layers.softmax(self.logits)
self.loss = layers.reduce_mean(layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss, global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(self.Yoh, self.preds, self.loss,
self.learning_rate)
def createModel(self):
print("CREATING MODEL")
self.global_step = tf.Variable(0, trainable=False)
self.is_training = tf.placeholder_with_default(True, [], name='is_training')
self.learning_rate = layers.decayLearningRate(LEARNING_RATE, self.global_step, DECAY_STEPS, DECAY_RATE)
self.X = tf.placeholder(dtype=tf.float32, shape=[None, self.dataset.frames, self.dataset.h, self.dataset.w, self.dataset.c])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
reuse = None
towersLogits = []
for sequence_image in range(self.dataset.frames):
net = self.vgg16(self.X[:, sequence_image], reuse)
towersLogits.append(net)
reuse = True
net = layers.stack(towersLogits)
del towersLogits[:]
net = layers.transpose(net, [1, 2, 3, 0, 4])
net = layers.reshape(net, [-1, net.shape[1], net.shape[2], net.shape[3] * net.shape[4]])
net = layers.fc(net, 512, name='fc5', weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.squeeze_and_excite2d(net, indexHeight=1, indexWidth=2, name='se5', filters=512)
net = layers.flatten(net, name='flatten')
net = layers.fc(net, 4096, name='fc6', weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.fc(net, 4096, name='fc7', weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
self.logits = layers.fc(net, self.dataset.num_classes, activation_fn=None, name='fc8',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
self.preds = layers.softmax(self.logits)
cross_entropy_loss = layers.reduce_mean(layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
regularization_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
self.loss = cross_entropy_loss + REGULARIZER_SCALE * tf.reduce_sum(regularization_loss)
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss, global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(self.Yoh, self.preds, self.loss,
self.learning_rate)
def modelTower(self):
self.is_training = tf.placeholder_with_default(True, [],
name='is_training')
self.towerImage = tf.placeholder(
dtype=tf.float32,
shape=[None, self.dataset.h, self.dataset.w, self.dataset.c])
bn_params = {
'decay': 0.999,
'center': True,
'scale': True,
'epsilon': 0.001,
'updates_collections': None,
'is_training': self.is_training
}
net = self.vgg(self.towerImage)
net_shape = net.get_shape()
net = tf.reshape(net, [
BATCH_SIZE,
int(net_shape[1]) * int(net_shape[2]) * int(net_shape[3])
])
self.towerNet = layers.fc(
net,
64,
name='spatial_FC',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE),
normalizer_fn=layers.batchNormalization,
normalizer_params=bn_params)
def modelLogits(self):
self.towerLogits = tf.placeholder(dtype=tf.float32, shape=[None, self.dataset.frames, 3, 3, 512])
net = self.towerLogits
net = layers.transpose(net, [0, 2, 3, 1, 4])
net = layers.reshape(net, [-1, net.shape[1], net.shape[2], net.shape[3] * net.shape[4]])
net = layers.fc(net, 512, name='fc5', weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.squeeze_and_excite2d(net, indexHeight=1, indexWidth=2, name='se5', filters=512)
net = layers.flatten(net, name='flatten')
net = layers.fc(net, 4096, name='fc6', weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.fc(net, 4096, name='fc7', weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
self.logits = layers.fc(net, self.dataset.num_classes, activation_fn=None, name='fc8',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
def createModelLowMemory(self):
print("CREATING MODEL")
self.global_step = tf.Variable(0, trainable=False)
self.is_training = tf.placeholder_with_default(True, [], name='is_training')
self.learning_rate = layers.decayLearningRate(LEARNING_RATE, self.global_step, DECAY_STEPS, DECAY_RATE)
self.X = tf.placeholder(dtype=tf.float32, shape=[None, self.dataset.frames, self.dataset.h, self.dataset.w, self.dataset.c])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
reuse = None
towersLogits = []
for sequence_image in range(self.dataset.frames):
net = self.mt_loop_low_memory(self.X[:, sequence_image], reuse)
towersLogits.append(net)
reuse = True
net = layers.stack(towersLogits)
del towersLogits[:]
net = layers.transpose(net, [1, 2, 3, 0, 4])
net = layers.reshape(net, [-1, net.shape[1], net.shape[2], net.shape[3] * net.shape[4]])
net = layers.fc(net, 512, name='fc5')
net = layers.flatten(net, name='flatten')
net = layers.fc(net, 4096, name='fc6')
net = layers.fc(net, 4096, name='fc7')
self.logits = layers.fc(net, self.dataset.num_classes, activation_fn=None, name='fc8')
self.preds = layers.softmax(self.logits)
self.loss = layers.reduce_mean(layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
self.opt = layers.sgd(self.learning_rate)
self.train_op = self.opt.minimize(self.loss, global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(self.Yoh, self.preds, self.loss,
self.learning_rate)
def modelTower(self):
self.is_training = tf.placeholder_with_default(True, [], name='is_training')
self.towerImage = tf.placeholder(dtype=tf.float32, shape=[None, self.dataset.h, self.dataset.w, self.dataset.c])
bn_params = {'decay': 0.999, 'center': True, 'scale': True, 'epsilon': 0.001, 'updates_collections': None,
'is_training': self.is_training}
net = self.vgg(self.towerImage)
net_shape = net.get_shape()
net = tf.reshape(net, [BATCH_SIZE, int(net_shape[1]) * int(net_shape[2]) * int(net_shape[3])])
self.towerNet = layers.fc(net, 64, name='spatial_FC',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE),
normalizer_fn=layers.batchNormalization, normalizer_params=bn_params)
def createModel(self):
print("CREATING MODEL")
self.global_step = tf.Variable(0, trainable=False)
self.is_training = tf.placeholder_with_default(True, [], name='is_training')
self.learning_rate = layers.decayLearningRate(LEARNING_RATE, self.global_step, DECAY_STEPS, DECAY_RATE)
self.X = tf.placeholder(dtype=tf.float32, shape=[None, self.dataset.frames, self.dataset.h, self.dataset.w, self.dataset.c])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
net = self.X
if net.shape[-1] > 1:
net = layers.rgb_to_grayscale(net)
if len(net.shape) > 4:
net = tf.transpose(net, [0, 2, 3, 1, 4])
net = tf.reshape(net, [-1, net.shape[1], net.shape[2], net.shape[3] * net.shape[4]])
bn_params = {'decay': 0.999, 'center': True, 'scale': True, 'epsilon': 0.001,
'updates_collections': None, 'is_training': self.is_training}
net = layers.conv2d(net, filters=96, kernel_size=3, padding='VALID', stride=2, name='conv1',
normalizer_fn=layers.batchNormalization, normalizer_params=bn_params,
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.max_pool2d(net, 3, 2, name='max_pool1')
net = layers.squeeze_and_excite2d(net, indexHeight=1, indexWidth=2, name='se1', filters=96)
net = layers.conv2d(net, filters=256, kernel_size=3, padding='VALID', stride=2, name='conv2',
normalizer_fn=layers.batchNormalization, normalizer_params=bn_params,
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.max_pool2d(net, 3, 2, name='max_pool2')
net = layers.squeeze_and_excite2d(net, indexHeight=1, indexWidth=2, name='se2', filters=256)
net = layers.conv2d(net, filters=512, kernel_size=3, padding='SAME', stride=1, name='conv3',
normalizer_fn=layers.batchNormalization, normalizer_params=bn_params,
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.squeeze_and_excite2d(net, indexHeight=1, indexWidth=2, name='se3', filters=512)
net = layers.conv2d(net, filters=512, kernel_size=3, padding='SAME', stride=1, name='conv4',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.squeeze_and_excite2d(net, indexHeight=1, indexWidth=2, name='se4', filters=512)
net = layers.conv2d(net, filters=512, kernel_size=3, padding='SAME', stride=1, name='conv5',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.max_pool2d(net, 3, 2, name='max_pool2')
net = layers.squeeze_and_excite2d(net, indexHeight=1, indexWidth=2, name='se5', filters=512)
net = layers.flatten(net, name='flatten')
net = layers.fc(net, 4096, name='fc6', weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.fc(net, 4096, name='fc7', weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
self.logits = layers.fc(net, self.dataset.num_classes, activation_fn=None, name='fc8', weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
self.preds = layers.softmax(self.logits)
cross_entropy_loss = layers.reduce_mean(layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
regularization_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
self.loss = cross_entropy_loss + REGULARIZER_SCALE * tf.reduce_sum(regularization_loss)
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss, global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(self.Yoh, self.preds, self.loss,
self.learning_rate)
def createModel(self):
print("CREATING MODEL")
self.global_step = tf.Variable(0, trainable=False)
self.is_training = tf.placeholder_with_default(True, [],
name='is_training')
self.learning_rate = layers.decayLearningRate(LEARNING_RATE,
self.global_step,
DECAY_STEPS, DECAY_RATE)
self.X = tf.placeholder(dtype=tf.float32,
shape=[
None, self.dataset.frames, self.dataset.h,
self.dataset.w, self.dataset.c
])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
net = self.X
conv3d_kernel = [3, 3, 3]
max3d_pool_kernel = [3, 3, 2]
bn_params = {
'decay': 0.999,
'center': True,
'scale': True,
'epsilon': 0.001,
'updates_collections': None,
'is_training': self.is_training
}
net = layers.transpose(net, [0, 2, 3, 1, 4])
net = layers.conv3d(
net,
filters=48,
kernel_size=conv3d_kernel,
padding='VALID',
stride=2,
name='conv1',
normalizer_fn=layers.batchNormalization,
normalizer_params=bn_params,
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.max_pool3d(net,
max3d_pool_kernel,
2,
padding='VALID',
name='max_pool1')
net = layers.squeeze_and_excite3d(net,
indexHeight=1,
indexWidth=2,
indexSeq=3,
name='se1',
filters=48)
net = layers.conv3d(
net,
filters=256,
kernel_size=conv3d_kernel,
padding='VALID',
stride=2,
name='conv2',
normalizer_fn=layers.batchNormalization,
normalizer_params=bn_params,
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.max_pool3d(net,
max3d_pool_kernel,
2,
padding='VALID',
name='max_pool2')
net = layers.squeeze_and_excite3d(net,
indexHeight=1,
indexWidth=2,
indexSeq=3,
name='se2',
filters=256)
net = layers.reshape(
net, [-1, net.shape[1], net.shape[2], net.shape[3] * net.shape[4]])
net = layers.conv2d(
net,
filters=512,
kernel_size=3,
padding='SAME',
stride=1,
name='conv3',
normalizer_fn=layers.batchNormalization,
normalizer_params=bn_params,
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.squeeze_and_excite2d(net,
indexHeight=1,
indexWidth=2,
name='se3',
filters=512)
net = layers.conv2d(
net,
filters=512,
kernel_size=3,
padding='SAME',
stride=1,
name='conv4',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.squeeze_and_excite2d(net,
indexHeight=1,
indexWidth=2,
name='se4',
filters=512)
net = layers.conv2d(
net,
filters=512,
kernel_size=3,
padding='SAME',
stride=1,
name='conv5',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.max_pool2d(net, 3, 1, padding='VALID', name='max_pool5')
net = layers.squeeze_and_excite2d(net,
indexHeight=1,
indexWidth=2,
name='se5',
filters=512)
net = layers.flatten(net, name='flatten')
net = layers.fc(
net,
4096,
name='fc6',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.fc(
net,
4096,
name='fc7',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
self.logits = layers.fc(
net,
self.dataset.num_classes,
activation_fn=None,
name='fc8',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
self.preds = layers.softmax(self.logits)
cross_entropy_loss = layers.reduce_mean(
layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
regularization_loss = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
self.loss = cross_entropy_loss + REGULARIZER_SCALE * tf.reduce_sum(
regularization_loss)
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss,
global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(
self.Yoh, self.preds, self.loss, self.learning_rate)
def createModel(self):
print("CREATING MODEL")
self.global_step = tf.Variable(0, trainable=False)
self.is_training = tf.placeholder_with_default(True, [],
name='is_training')
self.learning_rate = layers.decayLearningRate(LEARNING_RATE,
self.global_step,
DECAY_STEPS, DECAY_RATE)
self.X = tf.placeholder(dtype=tf.float32,
shape=[
None, self.dataset.frames, self.dataset.h,
self.dataset.w, self.dataset.c
])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
net = self.X
if net.shape[-1] > 1:
net = layers.rgb_to_grayscale(net)
if len(net.shape) > 4:
net = tf.transpose(net, [0, 2, 3, 1, 4])
net = tf.reshape(
net,
[-1, net.shape[1], net.shape[2], net.shape[3] * net.shape[4]])
net = self.build_vgg16(net)
if len(net.shape) > 4:
net = tf.transpose(net, [0, 2, 3, 1, 4])
net = tf.reshape(
net,
[-1, net.shape[1], net.shape[2], net.shape[3] * net.shape[4]])
net = layers.flatten(net, name='flatten')
net = layers.fc(
net,
4096,
name='fc6',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
net = layers.fc(
net,
4096,
name='fc7',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
self.logits = layers.fc(
net,
self.dataset.num_classes,
activation_fn=None,
name='fc8',
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE))
self.preds = layers.softmax(self.logits)
cross_entropy_loss = layers.reduce_mean(
layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
regularization_loss = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
self.loss = cross_entropy_loss + REGULARIZER_SCALE * tf.reduce_sum(
regularization_loss)
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss,
global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(
self.Yoh, self.preds, self.loss, self.learning_rate)
def createModel(self):
print("CREATING MODEL")
self.global_step = tf.Variable(0, trainable=False)
self.is_training = tf.placeholder_with_default(True, [],
name='is_training')
self.learning_rate = layers.decayLearningRate(LEARNING_RATE,
self.global_step,
DECAY_STEPS, DECAY_RATE)
self.X = tf.placeholder(dtype=tf.float32,
shape=[
None, self.dataset.frames, self.dataset.h,
self.dataset.w, self.dataset.c
])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
bn_params = {
'decay': 0.999,
'center': True,
'scale': True,
'epsilon': 0.001,
'updates_collections': None,
'is_training': self.is_training
}
concated = None
reuse = None
for sequence_image in range(self.dataset.frames):
net = self.vgg(self.X[:, sequence_image], reuse)
net_shape = net.get_shape()
net = tf.reshape(net, [
BATCH_SIZE,
int(net_shape[1]) * int(net_shape[2]) * int(net_shape[3])
])
net = layers.fc(
net,
64,
name='spatial_FC',
reuse=reuse,
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE),
normalizer_fn=layers.batchNormalization,
normalizer_params=bn_params)
if concated is None:
concated = tf.expand_dims(net, axis=1)
else:
concated = tf.concat(
[concated, tf.expand_dims(net, axis=1)], axis=1)
reuse = True
net = concated
net_shape = net.get_shape()
net = tf.reshape(
net,
[BATCH_SIZE, int(net_shape[1]) * int(net_shape[2])])
layer_num = 1
for fully_connected_num in [64]:
net = layers.fc(
net,
fully_connected_num,
name='temporal_FC{}'.format(layer_num),
weights_regularizer=layers.l2_regularizer(REGULARIZER_SCALE),
normalizer_fn=layers.batchNormalization,
normalizer_params=bn_params)
layer_num += 1
self.logits = layers.fc(net,
self.dataset.num_classes,
activation_fn=None,
name='logits')
self.preds = layers.softmax(self.logits)
cross_entropy_loss = layers.reduce_mean(
layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
regularization_loss = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
self.loss = cross_entropy_loss + REGULARIZER_SCALE * tf.reduce_sum(
regularization_loss)
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss,
global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(
self.Yoh, self.preds, self.loss, self.learning_rate)
def createModelLowMemory(self):
print("CREATING MODEL")
self.global_step = tf.Variable(0, trainable=False)
self.is_training = tf.placeholder_with_default(True, [],
name='is_training')
self.learning_rate = layers.decayLearningRate(LEARNING_RATE,
self.global_step,
DECAY_STEPS, DECAY_RATE)
self.X = tf.placeholder(dtype=tf.float32,
shape=[
None, self.dataset.frames, self.dataset.h,
self.dataset.w, self.dataset.c
])
self.Y = tf.placeholder(dtype=tf.int32, shape=[None])
self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes)
concated = None
reuse = None
for sequence_image in range(self.dataset.frames):
net = self.vggLowMemory(self.X[:, sequence_image], reuse)
net_shape = net.get_shape()
net = tf.reshape(net, [
BATCH_SIZE,
int(net_shape[1]) * int(net_shape[2]) * int(net_shape[3])
])
net = layers.fc(net, 64, name='spatial_FC', reuse=reuse)
if concated is None:
concated = tf.expand_dims(net, axis=1)
else:
concated = tf.concat(
[concated, tf.expand_dims(net, axis=1)], axis=1)
reuse = True
net = concated
net_shape = net.get_shape()
net = tf.reshape(
net,
[BATCH_SIZE, int(net_shape[1]) * int(net_shape[2])])
layer_num = 1
for fully_connected_num in [64]:
net = layers.fc(net,
fully_connected_num,
name='temporal_FC{}'.format(layer_num))
layer_num += 1
self.logits = layers.fc(net,
self.dataset.num_classes,
activation_fn=None,
name='logits')
self.preds = layers.softmax(self.logits)
self.loss = layers.reduce_mean(
layers.softmax_cross_entropy(logits=self.logits, labels=self.Yoh))
self.opt = layers.adam(self.learning_rate)
self.train_op = self.opt.minimize(self.loss,
global_step=self.global_step)
self.accuracy, self.precision, self.recall = self.createSummaries(
self.Yoh, self.preds, self.loss, self.learning_rate)