def __arg_scope(self, weight_decay=0.0005, data_format='NHWC'):
"""Defines the VGG arg scope.
Args:
weight_decay: The l2 regularization coefficient.
Returns:
An arg_scope.
"""
with slim.arg_scope([slim.conv2d, slim.fully_connected],
activation_fn=tf.nn.relu,
weights_regularizer=slim.l2_regularizer(weight_decay),
weights_initializer=tf.contrib.layers.xavier_initializer(),
biases_initializer=tf.zeros_initializer()):
with slim.arg_scope([slim.conv2d, slim.max_pool2d],
padding='SAME',
data_format=data_format):
with slim.arg_scope([custom_layers.pad2d,
custom_layers.l2_normalization,
custom_layers.channel_to_last],
data_format=data_format) as sc:
return sc
python类l2_regularizer()的实例源码
def encoder(self, x, embedding, reuse=None):
with tf.variable_scope("encoder", reuse=reuse):
with slim.arg_scope([slim.conv2d],
stride=1, activation_fn=tf.nn.elu, padding="SAME",
weights_initializer=tf.contrib.layers.variance_scaling_initializer(),
weights_regularizer=slim.l2_regularizer(5e-4),
bias_initializer=tf.zeros_initializer()):
x = slim.conv2d(x, embedding, 3)
for i in range(self.conv_repeat_num):
channel_num = embedding * (i + 1)
x = slim.repeat(x, 2, slim.conv2d, channel_num, 3)
if i < self.conv_repeat_num - 1:
# Is using stride pooling more better method than max pooling?
# or average pooling
# x = slim.conv2d(x, channel_num, kernel_size=3, stride=2) # sub-sampling
x = slim.avg_pool2d(x, kernel_size=2, stride=2)
# x = slim.max_pooling2d(x, 3, 2)
x = tf.reshape(x, [-1, np.prod([8, 8, channel_num])])
return x
def decoder(self, z, embedding, reuse=None):
with tf.variable_scope("decoder", reuse=reuse):
with slim.arg_scope([slim.conv2d, slim.fully_connected],
weights_initializer=tf.contrib.layers.variance_scaling_initializer(),
weights_regularizer=slim.l2_regularizer(5e-4),
bias_initializer=tf.zeros_initializer()):
with slim.arg_scope([slim.conv2d], padding="SAME",
activation_fn=tf.nn.elu, stride=1):
x = slim.fully_connected(z, 8 * 8 * embedding, activation_fn=None)
x = tf.reshape(x, [-1, 8, 8, embedding])
for i in range(self.conv_repeat_num):
x = slim.repeat(x, 2, slim.conv2d, embedding, 3)
if i < self.conv_repeat_num - 1:
x = resize_nn(x, 2) # NN up-sampling
x = slim.conv2d(x, 3, 3, activation_fn=None)
return x
def create_model(self, model_input, vocab_size, num_frames, **unused_params):
"""Creates a model which uses a logistic classifier over the average of the
frame-level features.
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]
max_frames = model_input.get_shape().as_list()[1]
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}
def sub_moe(self,
model_input,
vocab_size,
num_mixtures = None,
l2_penalty=1e-8,
scopename="",
**unused_params):
num_mixtures = num_mixtures or FLAGS.moe_num_mixtures
gate_activations = slim.fully_connected(
model_input,
vocab_size * (num_mixtures + 1),
activation_fn=None,
biases_initializer=None,
weights_regularizer=slim.l2_regularizer(l2_penalty),
scope="gates"+scopename)
expert_activations = slim.fully_connected(
model_input,
vocab_size * num_mixtures,
activation_fn=None,
weights_regularizer=slim.l2_regularizer(l2_penalty),
scope="experts"+scopename)
gating_distribution = tf.nn.softmax(tf.reshape(
gate_activations,
[-1, num_mixtures + 1])) # (Batch * #Labels) x (num_mixtures + 1)
expert_distribution = tf.nn.sigmoid(tf.reshape(
expert_activations,
[-1, num_mixtures])) # (Batch * #Labels) x num_mixtures
final_probabilities_by_class_and_batch = tf.reduce_sum(
gating_distribution[:, :num_mixtures] * expert_distribution, 1)
final_probabilities = tf.reshape(final_probabilities_by_class_and_batch,
[-1, vocab_size])
return model_input, final_probabilities
def create_model(self,
model_input,
vocab_size,
num_frames,
**unused_params):
shape = model_input.get_shape().as_list()
frames_sum = tf.reduce_sum(tf.abs(model_input),axis=2)
frames_true = tf.ones(tf.shape(frames_sum))
frames_false = tf.zeros(tf.shape(frames_sum))
frames_bool = tf.reshape(tf.where(tf.greater(frames_sum, frames_false), frames_true, frames_false),[-1,shape[1],1])
activation_1 = tf.reduce_max(model_input, axis=1)
activation_2 = tf.reduce_sum(model_input*frames_bool, axis=1)/(tf.reduce_sum(frames_bool, axis=1)+1e-6)
activation_3 = tf.reduce_min(model_input, axis=1)
model_input_1, final_probilities_1 = self.sub_moe(activation_1,vocab_size,scopename="_max")
model_input_2, final_probilities_2 = self.sub_moe(activation_2,vocab_size,scopename="_mean")
model_input_3, final_probilities_3 = self.sub_moe(activation_3,vocab_size,scopename="_min")
final_probilities = tf.stack((final_probilities_1,final_probilities_2,final_probilities_3),axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[shape[2], 3, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
activations = tf.stack((model_input_1, model_input_2, model_input_3), axis=2)
weight = tf.nn.softmax(tf.einsum("aij,ijk->ajk", activations, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters = [1024, 1024, 1024],
filter_sizes = [1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
lstm_size = FLAGS.lstm_cells
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = self.rnn(cnn_output,lstm_size, num_frames,sub_scope="rnn%d"%(layer+1))
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale,vocab_size,scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
num_frames = tf.maximum(num_frames//pool_size,1)
final_probilities = tf.stack(final_probilities,axis=1)
moe_inputs = tf.stack(moe_inputs,axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters = [1024, 1024, 1024],
filter_sizes = [1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def create_model(self, model_input, vocab_size, num_frames, distill_labels=None, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
lstm_size = FLAGS.lstm_cells
pool_size = 2
cnn_input = model_input
cnn_size = FLAGS.cnn_cells
num_filters = [cnn_size, cnn_size, cnn_size*2]
filter_sizes = [1, 2, 3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = self.rnn(cnn_output,lstm_size, num_frames,sub_scope="rnn%d"%(layer+1))
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale,vocab_size,distill_labels=distill_labels, scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
num_frames = tf.maximum(num_frames//pool_size,1)
final_probilities = tf.stack(final_probilities,axis=1)
moe_inputs = tf.stack(moe_inputs,axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, lstm_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters=[1024,1024,1024],
filter_sizes=[1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters=[1024,1024,1024],
filter_sizes=[1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
lstm_size = FLAGS.lstm_cells
pool_size = 2
cnn_input = model_input
num_filters = [256, 256, 512]
filter_sizes = [1, 2, 3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = self.rnn_glu(cnn_output, lstm_size, num_frames, sub_scope="rnn%d"%(layer+1))
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale, vocab_size, scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
num_frames = tf.maximum(num_frames//pool_size,1)
final_probilities = tf.stack(final_probilities, axis=1)
moe_inputs = tf.stack(moe_inputs, axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1, vocab_size])
result["predictions"] = tf.reduce_mean(final_probilities, axis=1)
return result
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
lstm_size = FLAGS.lstm_cells
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = LstmMultiscaleModel().cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = LstmMultiscaleModel().rnn(cnn_output,lstm_size, num_frames,sub_scope="rnn%d"%(layer+1))
moe_inputs.append(cnn_multiscale)
final_probility = LstmMultiscaleModel().sub_moe(cnn_multiscale,vocab_size,scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
num_frames = tf.maximum(num_frames//pool_size,1)
final_probilities = tf.stack(final_probilities, axis=1)
moe_inputs = tf.stack(moe_inputs, axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", tf.stop_gradient(moe_inputs), weight2d), dim=1)
result = {}
result["predictions"] = tf.reduce_sum(tf.stop_gradient(final_probilities)*weight, axis=1)
return result
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = 10
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
if layer < 3:
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
else:
cnn_input = cnn_output
cnn_output, num_t = self.kmax(cnn_input, num_filters=features_size, filter_sizes=num_extend, sub_scope="kmax")
cnn_input = tf.reshape(cnn_output,[-1,features_size])
final_probilities = self.sub_moe(cnn_input,vocab_size)
final_probilities = tf.reshape(final_probilities,[-1,num_extend,vocab_size])
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", cnn_output, weight2d), dim=1)
result = {}
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters = [1024, 1024, 1024],
filter_sizes = [1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = tf.reduce_max(cnn_output,axis=1)
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale,vocab_size,scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
final_probilities = tf.stack(final_probilities,axis=1)
moe_inputs = tf.stack(moe_inputs,axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters = [1024, 1024, 1024],
filter_sizes = [1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
sub_bias = tf.get_variable(sub_scope+"cnn-bias-len%d"%fs,
shape=[nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter) + sub_bias)
cnn_output = tf.concat(cnn_outputs, axis=2)
cnn_output = slim.batch_norm(
cnn_output,
center=True,
scale=True,
is_training=FLAGS.train,
scope=sub_scope+"cluster_bn")
return cnn_output, max_frames
def create_model(self, model_input, vocab_size, num_frames, l2_penalty=1e-8, **unused_params):
num_extend = FLAGS.moe_num_extend
num_layers = num_extend
pool_size=2
cnn_input = model_input
num_filters=[256,256,512]
filter_sizes=[1,2,3]
features_size = sum(num_filters)
final_probilities = []
moe_inputs = []
for layer in range(num_layers):
cnn_output, num_t = self.cnn(cnn_input, num_filters=num_filters, filter_sizes=filter_sizes, sub_scope="cnn%d"%(layer+1))
cnn_output = tf.nn.relu(cnn_output)
cnn_multiscale = tf.reduce_max(cnn_output,axis=1)
moe_inputs.append(cnn_multiscale)
final_probility = self.sub_moe(cnn_multiscale,vocab_size,scopename="moe%d"%(layer+1))
final_probilities.append(final_probility)
num_t = pool_size*(num_t//pool_size)
cnn_output = tf.reshape(cnn_output[:,:num_t,:],[-1,num_t//pool_size,pool_size,features_size])
cnn_input = tf.reduce_max(cnn_output, axis=2)
final_probilities = tf.stack(final_probilities,axis=1)
moe_inputs = tf.stack(moe_inputs,axis=1)
weight2d = tf.get_variable("ensemble_weight2d",
shape=[num_extend, features_size, vocab_size],
regularizer=slim.l2_regularizer(1.0e-8))
weight = tf.nn.softmax(tf.einsum("aij,ijk->aik", moe_inputs, weight2d), dim=1)
result = {}
result["prediction_frames"] = tf.reshape(final_probilities,[-1,vocab_size])
result["predictions"] = tf.reduce_sum(final_probilities*weight,axis=1)
return result
def cnn(self,
model_input,
l2_penalty=1e-8,
num_filters=[1024, 1024, 1024],
filter_sizes=[1,2,3],
sub_scope="",
**unused_params):
max_frames = model_input.get_shape().as_list()[1]
num_features = model_input.get_shape().as_list()[2]
shift_inputs = []
for i in range(max(filter_sizes)):
if i == 0:
shift_inputs.append(model_input)
else:
shift_inputs.append(tf.pad(model_input, paddings=[[0,0],[i,0],[0,0]])[:,:max_frames,:])
cnn_outputs = []
for nf, fs in zip(num_filters, filter_sizes):
sub_input = tf.concat(shift_inputs[:fs], axis=2)
sub_filter = tf.get_variable(sub_scope+"cnn-filter-len%d"%fs,
shape=[num_features*fs, nf], dtype=tf.float32,
initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1),
regularizer=tf.contrib.layers.l2_regularizer(l2_penalty))
cnn_outputs.append(tf.einsum("ijk,kl->ijl", sub_input, sub_filter))
cnn_output = tf.concat(cnn_outputs, axis=2)
return cnn_output, max_frames
