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 modelOpt(self): self.global_step = tf.Variable(0, trainable=False) self.learning_rate = layers.decayLearningRate(LEARNING_RATE, self.global_step, DECAY_STEPS, DECAY_RATE) self.optLogits = tf.placeholder(dtype=tf.float32, shape=[None, self.dataset.num_classes]) self.Y = tf.placeholder(dtype=tf.int32, shape=[None]) self.Yoh = layers.toOneHot(self.Y, self.dataset.num_classes) self.preds = layers.softmax(self.optLogits) cross_entropy_loss = layers.reduce_mean(layers.softmax_cross_entropy(logits=self.optLogits, 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]]) 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)