class Softmax(Activation):
_expected_attributes = [
ChoiceAttribute('implementation', ['latency', 'stable', 'legacy'], default='stable')
]
def initialize(self):
super(Softmax, self).initialize()
class Pooling2D(Layer):
_expected_attributes = [
Attribute('in_height'),
Attribute('in_width'),
Attribute('out_height'),
Attribute('out_width'),
Attribute('n_filt'),
Attribute('pool_height'),
Attribute('pool_width'),
Attribute('stride_height'),
Attribute('stride_width'),
Attribute('pad_top'),
Attribute('pad_bottom'),
Attribute('pad_left'),
Attribute('pad_right'),
ChoiceAttribute('pool_op', ['Max', 'Average'])
]
def initialize(self):
if self.get_attr('data_format') == 'channels_last':
shape = [self.attributes['out_height'], self.attributes['out_width'], self.attributes['n_filt']]
dims = ['OUT_HEIGHT_{}'.format(self.index), 'OUT_WIDTH_{}'.format(self.index), 'N_FILT_{}'.format(self.index)]
else:
shape = [self.attributes['n_filt'], self.attributes['out_height'], self.attributes['out_width']]
dims = ['N_FILT_{}'.format(self.index), 'OUT_HEIGHT_{}'.format(self.index), 'OUT_WIDTH_{}'.format(self.index)]
self.add_output_variable(shape, dims)
self.set_attr('pool_op', self.get_attr('class_name').split('Pooling')[0])
self.set_attr('implementation', self.model.config.get_conv_implementation(self).lower())
class GRU(Layer):
_expected_attributes = [
Attribute('n_out'),
Attribute('activation', value_type=str),
Attribute('recurrent_activation', value_type=str),
Attribute('return_sequences', value_type=bool, default=False),
Attribute('return_state', value_type=bool, default=False),
ChoiceAttribute('direction', ['forward', 'backward'], default='forward'),
Attribute('time_major', value_type=bool, default=False),
ChoiceAttribute('apply_reset_gate', ['before', 'after'], default='after'),
WeightAttribute('weight'),
WeightAttribute('bias'),
WeightAttribute('recurrent_weight'),
TypeAttribute('weight'),
TypeAttribute('bias'),
TypeAttribute('recurrent_weight'),
]
def initialize(self):
if self.attributes['return_sequences']:
shape = [self.attributes['n_timesteps'], self.attributes['n_out']]
dims = ['N_TIME_STEPS_{}'.format(self.index), 'N_OUT_{}'.format(self.index)]
else:
shape = [self.attributes['n_out']]
dims = ['N_OUT_{}'.format(self.index)]
self.add_output_variable(shape, dims)
if self.attributes['return_state']:
state_shape = [self.attributes['n_out']]
state_dims = ['N_OUT_{}'.format(self.index)]
self.add_output_variable(state_shape, state_dims, out_name=self.outputs[1], var_name='layer{index}_h', type_name='layer{index}_h_t')
self.add_output_variable(state_shape, state_dims, out_name=self.outputs[2], var_name='layer{index}_c', type_name='layer{index}_c_t')
self.add_weights()
self.add_bias()
recurrent_weight = self.model.get_weights_data(self.name, 'recurrent_kernel')
self.add_weights_variable(name='recurrent_weight', var_name='wr{index}', data=recurrent_weight)
class GlobalPooling1D(Layer):
_expected_attributes = [
Attribute('n_in'),
Attribute('n_filt'),
ChoiceAttribute('pool_op', ['Max', 'Average'])
]
def initialize(self):
shape = [self.attributes['n_filt']]
dims = ['N_FILT_{}'.format(self.index)]
self.add_output_variable(shape, dims)
self.set_attr('pool_op', self.get_attr('class_name').split('Pooling')[0].replace('Global', ''))