def single_layer(self, model_layer, name):
""" build a single layer"""
# input layer
if model_layer['layer'] == 'input':
with tf.name_scope('input') as scope:
input_shape = str(model_layer['input_shape'])
inputs = utils.placeholder(shape=model_layer['input_shape'], name=name)
self.network[name] = layers.InputLayer(inputs)
self.placeholders[name] = inputs
self.feed_dict[name] = []
# dense layer
elif model_layer['layer'] == 'dense':
with tf.name_scope('dense') as scope:
if 'W' not in model_layer.keys():
model_layer['W'] = init.GlorotUniform(**self.seed)
self.network[name] = layers.DenseLayer(self.network[self.last_layer], num_units=model_layer['num_units'],
W=model_layer['W'],
b=None)
# convolution layer
elif (model_layer['layer'] == 'conv2d'):
with tf.name_scope('conv2d') as scope:
if 'W' not in model_layer.keys():
W = init.GlorotUniform(**self.seed)
else:
W = model_layer['W']
if 'padding' not in model_layer.keys():
padding = 'VALID'
else:
padding = model_layer['padding']
if 'strides' not in model_layer.keys():
strides = (1, 1)
else:
strides = model_layer['strides']
self.network[name] = layers.Conv2DLayer(self.network[self.last_layer], num_filters=model_layer['num_filters'],
filter_size=model_layer['filter_size'],
W=W,
padding=padding,
strides=strides)
elif model_layer['layer'] == 'conv1d':
with tf.name_scope('conv1d') as scope:
if 'W' not in model_layer.keys():
W = init.GlorotUniform(**self.seed)
else:
W = model_layer['W']
if 'padding' not in model_layer.keys():
padding = 'VALID'
else:
padding = model_layer['padding']
if 'strides' not in model_layer.keys():
strides = 1
else:
strides = model_layer['strides']
reverse=False
if 'reverse' in model_layer:
reverse = model_layer['reverse']
self.network[name] = layers.Conv1DLayer(self.network[self.last_layer], num_filters=model_layer['num_filters'],
filter_size=model_layer['filter_size'],
W=W,
padding=padding,
strides=strides,
reverse=reverse)
# convolution layer
elif (model_layer['layer'] == 'conv2d_transpose'):
if 'W' not in model_layer.keys():
W = init.GlorotUniform(**self.seed)
else:
W = model_layer['W']
if 'padding' not in model_layer.keys():
padding = 'SAME'
else:
padding = model_layer['padding']
if 'strides' not in model_layer.keys():
strides = (1, 1)
else:
strides = model_layer['strides']
self.network[name] = layers.TransposeConv2DLayer(self.network[self.last_layer], num_filters=model_layer['num_filters'],
filter_size=model_layer['filter_size'],
W=W,
padding=padding,
strides=strides)
elif model_layer['layer'] == 'conv1d_transpose':
if 'W' not in model_layer.keys():
W = init.GlorotUniform(**self.seed)
else:
W = model_layer['W']
if 'padding' not in model_layer.keys():
padding = 'SAME'
else:
padding = model_layer['padding']
if 'strides' not in model_layer.keys():
strides = 1
else:
strides = model_layer['strides']
self.network[name] = layers.TransposeConv1DLayer(self.network[self.last_layer], num_filters=model_layer['num_filters'],
filter_size=model_layer['filter_size'],
W=W,
padding=padding,
strides=strides)
# concat layer
elif model_layer['layer'] == 'concat':
self.network[name] = layers.ConcatLayer([self.network[self.last_layer], self.network[model_layer['concat']]])
# element-size sum layer
elif model_layer['layer'] == 'sum':
self.network[name] = layers.ElemwiseSumLayer([self.network[self.last_layer], model_layer['sum']])
# reshape layer
elif model_layer['layer'] == 'reshape':
self.network[name] = layers.ReshapeLayer(self.network[self.last_layer], model_layer['reshape'])
elif model_layer['layer'] == 'reduce_max':
self.network[name] = layers.MaxLayer(self.network[self.last_layer], axis=1)
elif model_layer['layer'] == 'reduce_mean':
self.network[name] = layers.MeanLayer(self.network[self.last_layer], axis=1)
elif model_layer['layer'] == 'softmax2D':
self.network[name] = layers.Softmax2DLayer(self.network[self.last_layer])
self.last_layer = name
def build_layers(self, model_layers, supervised=True):
self.network = OrderedDict()
name_gen = NameGenerator()
self.num_dropout = 0
self.num_inputs = 0
self.last_layer = ''
# loop to build each layer of network
for model_layer in model_layers:
layer = model_layer['layer']
# name of layer
if 'name' in model_layer:
name = model_layer['name']
else:
name = name_gen.generate_name(layer)
# set scope for each layer
with tf.name_scope(name) as scope:
if layer == "input":
# add input layer
self.single_layer(model_layer, name)
elif layer == 'embedding':
vocab_size = model_layer['vocab_size']
embedding_size = model_layer['embedding_size']
if 'max_norm' in model_layer:
max_norm = model_layer['max_norm']
else:
max_norm = None
self.network[name] = layers.EmbeddingLayer(self.network[self.last_layer], vocab_size, embedding_size, max_norm)
self.last_layer = name
elif (layer == 'variational') | (layer == 'variational_normal'):
if 'name' in model_layer:
name = model_layer['name']
else:
name = 'Z'
self.network[name+'_mu'] = layers.DenseLayer(self.network[self.last_layer], num_units=model_layer['num_units'], b=init.GlorotUniform(), **self.seed)
self.network[name+'_logvar'] = layers.DenseLayer(self.network[self.last_layer], num_units=model_layer['num_units'], b=init.GlorotUniform(), **self.seed)
self.network[name+'_sample'] = layers.VariationalSampleLayer(self.network[name+'_mu'], self.network[name+'_logvar'])
self.last_layer = name+'_sample'
elif layer == 'variational_softmax':
if 'hard' in model_layer:
hard = model_layer['hard']
else:
hard = False
num_categories, num_classes = model_layer['shape']
if 'temperature' in model_layer:
temperature = model_layer['temperature']
else:
temperature = 5.0
self.feed_dict['temperature'] = temperature
self.placeholders['temperature'] = tf.placeholder(dtype=tf.float32, name="temperature")
if 'name' in model_layer:
name = model_layer['name']
else:
name = 'Z'
if 'output' in model_layer:
output = model_layer['output']
else:
output = 'softmax'
self.network[name+'_logits'] = layers.DenseLayer(self.network[self.last_layer], num_units=num_categories*num_classes, b=init.GlorotUniform())
self.network[name+'_logits_reshape'] = layers.ReshapeLayer(self.network[name+'_logits'], shape=[-1, num_categories, num_classes])
self.network[name+'_sample'] = layers.CategoricalSampleLayer(self.network[name+'_logits_reshape'], temperature=temperature, hard=hard)
self.network[name+'_softmax'] = layers.Softmax2DLayer(self.network[name+'_logits_reshape'])
if output == 'hard':
self.network[name] = layers.ReshapeLayer(self.network[name+'_sample'], shape=[-1, num_categories*num_classes])
else:
self.network[name] = layers.ReshapeLayer(self.network[name+'_softmax'], shape=[-1, num_categories*num_classes])
self.last_layer = name
else:
if layer == 'conv1d_residual':
self.conv1d_residual_block(model_layer, name)
elif layer == 'conv2d_residual':
self.conv2d_residual_block(model_layer, name)
elif layer == 'dense_residual':
self.dense_residual_block(model_layer, name)
else:
# add core layer
self.single_layer(model_layer, name)
# add Batch normalization layer
if 'norm' in model_layer:
if 'batch' in model_layer['norm']:
with tf.name_scope("norm") as scope:
new_layer = name + '_batch' #str(counter) + '_' + name + '_batch'
self.network[new_layer] = layers.BatchNormLayer(self.network[self.last_layer], self.placeholders['is_training'])
self.last_layer = new_layer
else:
if (model_layer['layer'] == 'dense') | (model_layer['layer'] == 'conv1d') | (model_layer['layer'] == 'conv2d'):
if 'b' in model_layer:
if model_layer['b'] != None:
with tf.name_scope("bias") as scope:
b = init.Constant(model_layer['b'])
new_layer = name+'_bias'
self.network[new_layer] = layers.BiasLayer(self.network[self.last_layer], b=b)
self.last_layer = new_layer
elif 'norm' not in model_layer:
with tf.name_scope("bias") as scope:
b = init.GlorotUniform()
new_layer = name+'_bias'
self.network[new_layer] = layers.BiasLayer(self.network[self.last_layer], b=b)
self.last_layer = new_layer
# add activation layer
if 'activation' in model_layer:
new_layer = name+'_active'
self.network[new_layer] = layers.ActivationLayer(self.network[self.last_layer], function=model_layer['activation'], name=scope)
self.last_layer = new_layer
# add max-pooling layer
if 'max_pool' in model_layer:
new_layer = name+'_maxpool' # str(counter) + '_' + name+'_pool'
if 'max_pool_strides' in model_layer:
strides = model_layer['max_pool_strides']
else:
strides = model_layer['max_pool']
if isinstance(model_layer['max_pool'], (tuple, list)):
self.network[new_layer] = layers.MaxPool2DLayer(self.network[self.last_layer], pool_size=model_layer['max_pool'], strides=strides, name=name+'_maxpool')
else:
self.network[new_layer] = layers.MaxPool1DLayer(self.network[self.last_layer], pool_size=model_layer['max_pool'], strides=strides, name=name+'_maxpool')
self.last_layer = new_layer
# add mean-pooling layer
elif 'mean_pool' in model_layer:
new_layer = name+'_meanpool' # str(counter) + '_' + name+'_pool'
if 'mean_pool_strides' in model_layer:
strides = model_layer['mean_pool_strides']
else:
strides = model_layer['mean_pool']
if isinstance(model_layer['mean_pool'], (tuple, list)):
self.network[new_layer] = layers.MeanPool2DLayer(self.network[self.last_layer], pool_size=model_layer['mean_pool'], strides=strides, name=name+'_meanpool')
else:
self.network[new_layer] = layers.MeanPool1DLayer(self.network[self.last_layer], pool_size=model_layer['mean_pool'], strides=strides, name=name+'_meanpool')
self.last_layer = new_layer
# add global-pooling layer
elif 'global_pool' in model_layer:
new_layer = name+'_globalpool'
self.network[new_layer] = layers.GlobalPoolLayer(self.network[self.last_layer], func=model_layer['global_pool'], name=name+'_globalpool')
self.last_layer = new_layer
# add dropout layer
if 'dropout' in model_layer:
new_layer = name+'_dropout' # str(counter) + '_' + name+'_dropout'
placeholder_name = 'keep_prob_'+str(self.num_dropout)
self.placeholders[placeholder_name] = tf.placeholder(tf.float32, name=placeholder_name)
self.feed_dict[placeholder_name] = 1-model_layer['dropout']
self.num_dropout += 1
self.network[new_layer] = layers.DropoutLayer(self.network[self.last_layer], keep_prob=self.placeholders[placeholder_name], name=name+'_dropout')
self.last_layer = new_layer
if ('reshape' in model_layer) & (layer != 'reshape'):
new_layer = name+'_reshape'
self.network[new_layer] = layers.ReshapeLayer(self.network[self.last_layer], model_layer['reshape'])
self.last_layer = new_layer
if supervised:
self.network['output'] = self.network.pop(self.last_layer)
shape = self.network['output'].get_output_shape()
targets = utils.placeholder(shape=shape, name='output')
self.placeholders['targets'] = targets
self.feed_dict['targets'] = []
else:
self.network['X'] = self.network.pop(self.last_layer)
self.placeholders['targets'] = self.placeholders['inputs']
self.feed_dict['targets'] = []
return self.network, self.placeholders, self.feed_dict
def build_layers(self, model_layers, supervised=True, use_scope=False):
self.network = OrderedDict()
name_gen = NameGenerator()
self.num_dropout = 0
self.num_inputs = 0
self.last_layer = ''
# loop to build each layer of network
for model_layer in model_layers:
layer = model_layer['layer']
# name of layer
if 'name' in model_layer:
name = model_layer['name']
else:
name = name_gen.generate_name(layer)
if layer == "input":
# add input layer
self.single_layer(model_layer, name)
else:
if layer == 'conv1d_residual':
self.conv1d_residual_block(model_layer, name)
elif layer == 'conv2d_residual':
self.conv2d_residual_block(model_layer, name)
elif layer == 'dense_residual':
self.dense_residual_block(model_layer, name)
elif layer == 'variational':
self.network['encode_mu'] = layers.DenseLayer(
self.network[self.last_layer],
num_units=model_layer['num_units'],
**self.seed)
self.network['encode_logsigma'] = layers.DenseLayer(
self.network[self.last_layer],
num_units=model_layer['num_units'],
**self.seed)
self.network['Z'] = layers.VariationalSampleLayer(
self.network['encode_mu'],
self.network['encode_logsigma'])
self.last_layer = 'Z'
else:
# add core layer
self.single_layer(model_layer, name)
# add Batch normalization layer
if 'norm' in model_layer:
if 'batch' in model_layer['norm']:
new_layer = name + '_batch' #str(counter) + '_' + name + '_batch'
self.network[new_layer] = layers.BatchNormLayer(
self.network[self.last_layer],
self.placeholders['is_training'])
self.last_layer = new_layer
else:
if (model_layer['layer']
== 'dense') | (model_layer['layer'] == 'conv1d') | (
model_layer['layer'] == 'conv2d'):
if 'b' in model_layer:
if model_layer['b'] != None:
b = init.Constant(0.05)
new_layer = name + '_bias'
self.network[new_layer] = layers.BiasLayer(
self.network[self.last_layer], b=b)
self.last_layer = new_layer
elif 'norm' not in model_layer:
b = init.Constant(0.05)
new_layer = name + '_bias'
self.network[new_layer] = layers.BiasLayer(
self.network[self.last_layer], b=b)
self.last_layer = new_layer
else:
if 'b' in model_layer:
if model_layer['b'] != None:
b = init.Constant(0.05)
new_layer = name + '_bias'
self.network[new_layer] = layers.BiasLayer(
self.network[self.last_layer], b=b)
self.last_layer = new_layer
# add activation layer
if 'activation' in model_layer:
new_layer = name + '_active'
self.network[new_layer] = layers.ActivationLayer(
self.network[self.last_layer],
function=model_layer['activation'])
self.last_layer = new_layer
# add max-pooling layer ### Modified this from the older pool_size
if 'max_pool' in model_layer:
new_layer = name + '_maxpool' # str(counter) + '_' + name+'_pool'
if len(self.network[self.last_layer].output_shape) == 4:
if isinstance(model_layer['max_pool'], (tuple, list)):
self.network[new_layer] = layers.MaxPool2DLayer(
self.network[self.last_layer],
pool_size=model_layer['max_pool'])
else:
self.network[new_layer] = layers.MaxPool2DLayer(
self.network[self.last_layer],
pool_size=(model_layer['max_pool'], 1))
self.last_layer = new_layer
# add mean-pooling layer
elif 'mean_pool' in model_layer:
new_layer = name + '_meanpool' # str(counter) + '_' + name+'_pool'
if isinstance(model_layer['mean_pool'], (tuple, list)):
self.network[new_layer] = layers.MeanPool2DLayer(
self.network[self.last_layer],
pool_size=model_layer['mean_pool'],
name=name + '_meanpool')
else:
self.network[new_layer] = layers.MeanPool1DLayer(
self.network[self.last_layer],
pool_size=model_layer['mean_pool'],
name=name + '_meanpool')
self.last_layer = new_layer
# add global-pooling layer
elif 'global_pool' in model_layer:
print('global_pool')
new_layer = name + '_globalpool'
self.network[new_layer] = layers.GlobalPoolLayer(
self.network[self.last_layer],
func=model_layer['global_pool'],
name=name + '_globalpool')
self.last_layer = new_layer
# add dropout layer
if 'dropout' in model_layer:
new_layer = name + '_dropout' # str(counter) + '_' + name+'_dropout'
placeholder_name = 'keep_prob_' + str(self.num_dropout)
self.placeholders[placeholder_name] = tf.placeholder(
tf.float32, name=placeholder_name)
self.feed_dict[placeholder_name] = 1 - model_layer['dropout']
self.num_dropout += 1
self.network[new_layer] = layers.DropoutLayer(
self.network[self.last_layer],
keep_prob=self.placeholders[placeholder_name])
self.last_layer = new_layer
if supervised:
self.network['output'] = self.network.pop(self.last_layer)
shape = self.network['output'].get_output_shape()
targets = utils.placeholder(shape=shape, name='output')
self.placeholders['targets'] = targets
self.feed_dict['targets'] = []
else:
self.network['X'] = self.network.pop(self.last_layer)
self.placeholders['targets'] = self.placeholders['inputs'][0]
self.feed_dict['targets'] = []
self.feed_dict['KL_weight'] = 1.0
self.placeholders['KL_weight'] = tf.placeholder(tf.float32)
return self.network, self.placeholders, self.feed_dict
