def BL(template, dropout=0.1, regularizer=None, constraint=None):
"""
inputs
----
template: a list of network dimensions including input and output, e.g., [[40,10], [120,5], [3,1]]
dropout: dropout percentage
regularizer: keras regularizer object
constraint: keras constraint object
outputs
------
keras model object
"""
inputs = keras.layers.Input(template[0])
x = inputs
for k in range(1, len(template) - 1):
x = Layers.BL(template[k], regularizer, constraint)(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Dropout(dropout)(x)
x = Layers.BL(template[-1], regularizer, constraint)(x)
outputs = keras.layers.Activation('softmax')(x)
model = keras.Model(inputs=inputs, outputs=outputs)
optimizer = keras.optimizers.Adam(0.01)
model.compile(optimizer, 'categorical_crossentropy', [
'acc',
])
return model
def get_nonlinear_encoder(inputs,
input_shape,
target_shape,
prefix=None,
complexity='low',
conv_regularizer=None,
conv_constraint=None,
bl_regularizer=None,
bl_constraint=None):
h_scale = np.log2(np.floor(float(input_shape[0]) / target_shape[0]))
w_scale = np.log2(np.floor(float(input_shape[1]) / target_shape[1]))
# number of upsampling step
nb_downsampling = max(1, int(min(h_scale, w_scale)))
if complexity == 'low':
nb_conv_per_block = 3 if target_shape[0] == target_shape[
1] and target_shape[0] == 6 else 4
else:
nb_conv_per_block = 4 if target_shape[0] == target_shape[
1] and target_shape[0] == 6 else 3
nb_filter1 = 32
nb_filter2 = 64
for down_iter in range(nb_downsampling):
suffix = '_' + str(down_iter)
inputs = downsampling_block(inputs, nb_filter1, nb_filter2,
nb_conv_per_block, prefix, suffix,
conv_regularizer, conv_constraint)
nb_filter1 = nb_filter2
nb_filter2 = min(nb_filter2 * 2, 256)
if nb_downsampling == 1 and complexity == 'high':
inputs = downsampling_block(inputs, nb_filter1, nb_filter2,
nb_conv_per_block, prefix, '_1',
conv_regularizer, conv_constraint, False)
name = 'nonlinear_encoder_bl' if prefix is None else prefix + 'nonlinear_encoder_bl'
outputs = Layers.BL(target_shape,
kernel_regularizer=bl_regularizer,
kernel_constraint=bl_constraint,
name=name)(inputs)
return outputs
def TABL(template,
dropout=0.1,
projection_regularizer=None,
projection_constraint=None,
attention_regularizer=None,
attention_constraint=None):
"""
Temporal Attention augmented Bilinear Layer network, refer to the paper https://arxiv.org/abs/1712.00975
inputs
----
template: a list of network dimensions including input and output, e.g., [[40,10], [120,5], [3,1]]
dropout: dropout percentage
projection_regularizer: keras regularizer object for projection matrices
projection_constraint: keras constraint object for projection matrices
attention_regularizer: keras regularizer object for attention matrices
attention_constraint: keras constraint object for attention matrices
outputs
------
keras model object
"""
inputs = keras.layers.Input(template[0])
x = inputs
for k in range(1, len(template) - 1):
x = Layers.BL(template[k], projection_regularizer,
projection_constraint)(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Dropout(dropout)(x)
x = Layers.TABL(template[-1], projection_regularizer,
projection_constraint, attention_regularizer,
attention_constraint)(x)
outputs = keras.layers.Activation('softmax')(x)
model = keras.Model(inputs=inputs, outputs=outputs)
optimizer = keras.optimizers.Adam(0.01)
model.compile(optimizer, 'categorical_crossentropy', [
'acc',
])
return model
