def build(self, hp, inputs=None):
input_node = nest.flatten(inputs)[0]
output_node = input_node
vectorizer = self.vectorizer or hp.Choice(
'vectorizer', ['sequence', 'ngram'], default='sequence')
if not isinstance(input_node, node_module.TextNode):
raise ValueError('The input_node should be a TextNode.')
if vectorizer == 'ngram':
output_node = preprocessor_module.TextToNgramVector()(output_node)
output_node = block_module.DenseBlock()(output_node)
else:
output_node = preprocessor_module.TextToIntSequence()(output_node)
output_node = block_module.EmbeddingBlock(
pretraining=self.pretraining)(output_node)
output_node = block_module.ConvBlock(separable=True)(output_node)
output_node = block_module.SpatialReduction()(output_node)
output_node = block_module.DenseBlock()(output_node)
return output_node
def test_dense_block():
input_shape = (32, )
block = block_module.DenseBlock()
hp = kerastuner.HyperParameters()
block.build(hp, ak.Input(shape=input_shape).build())
assert name_in_hps('num_layers', hp)
assert name_in_hps('use_batchnorm', hp)
def assemble(self, input_node):
output_node = input_node
ratio = self.sw_ratio()
if not isinstance(input_node, node.TextNode):
raise ValueError('The input_node should be a TextNode.')
if ratio < 1500:
output_node = processor.TextToNgramVector()(output_node)
output_node = block.DenseBlock()(output_node)
else:
output_node = processor.TextToIntSequence()(output_node)
output_node = block.EmbeddingBlock(
pretrained=(ratio < 15000))(output_node)
output_node = block.ConvBlock(separable=True)(output_node)
return output_node
def build_body(self, hp, input_node):
if len(self.heads) > 1:
module_type_choices = ['dense']
else:
module_type_choices = ['lightgbm', 'dense']
module_type = self.module_type or hp.Choice(
'module_type', module_type_choices, default=module_type_choices[0])
if module_type == 'dense':
output_node = block.DenseBlock()(input_node)
elif module_type == 'lightgbm':
output_node = preprocessor.LightGBMBlock(
seed=self.seed)(input_node)
else:
raise ValueError(
'Unsupported module'
'type: {module_type}'.format(module_type=module_type))
nest.flatten(output_node)[0].shape = self.output_shape
return output_node
def build_body(self, hp, input_node):
if self.num_heads > 1:
block_type = ['dense']
else:
block_type = ['lightgbm', 'dense']
block_type = self.block_type or hp.Choice('block_type',
block_type,
default=block_type[0])
if block_type == 'dense':
output_node = block_module.DenseBlock()(input_node)
elif block_type == 'lightgbm':
output_node = preprocessor_module.LightGBM(
seed=self.seed)(input_node)
else:
raise ValueError('Expect the block_type to be "dense" or "lightgbm", '
'but got {block_type}'.format(block_type=block_type))
nest.flatten(output_node)[0].shape = self.output_shape
return output_node
def assemble(self, input_node):
block = hyperblock.ImageBlock(seed=self.seed)
if max(self._shape[0], self._shape[1]) < 32:
if self._num_samples < 10000:
self.hps.append(hp_module.Choice(
block.name + '_resnet/v1/conv4_depth', [6],
default=6))
self.hps.append(hp_module.Choice(
block.name + '_resnet/v2/conv4_depth', [6],
default=6))
self.hps.append(hp_module.Choice(
block.name + '_resnet/next/conv4_depth', [6],
default=6))
self.hps.append(hp_module.Int(
block.name + '_xception/num_residual_blocks', 2, 4,
default=4))
output_node = block(input_node)
output_node = block_module.SpatialReduction()(output_node)
return block_module.DenseBlock()(output_node)
