def __init__(self, component):
"""Initializes layers.
Args:
component: Parent ComponentBuilderBase object.
"""
layers = [
network_units.Layer(self, 'lengths', -1),
network_units.Layer(self, 'scores', -1),
network_units.Layer(self, 'logits', -1),
network_units.Layer(self, 'arcs', -1),
]
super(MstSolverNetwork, self).__init__(component, init_layers=layers)
self._attrs = network_units.get_attrs_with_defaults(
component.spec.network_unit.parameters,
defaults={
'forest': False,
'loss': 'softmax',
'crf_max_dynamic_range': 20,
})
check.Eq(len(self._fixed_feature_dims.items()), 0,
'Expected no fixed features')
check.Eq(len(self._linked_feature_dims.items()), 2,
'Expected two linked features')
check.In('lengths', self._linked_feature_dims,
'Missing required linked feature')
check.In('scores', self._linked_feature_dims,
'Missing required linked feature')
def testCheckIn(self):
check.In('a', ('a', 'b', 'c'), 'foo')
check.In('b', {'a': 1, 'b': 2}, 'bar')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.In('d', ('a', 'b', 'c'), 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.In('c', {'a': 1, 'b': 2}, 'baz', RuntimeError)
def __init__(self, component):
"""Initializes weights and layers.
Args:
component: Parent ComponentBuilderBase object.
"""
super(BiaffineLabelNetwork, self).__init__(component)
parameters = component.spec.network_unit.parameters
self._num_labels = int(parameters['num_labels'])
check.Gt(self._num_labels, 0, 'Expected some labels')
check.Eq(len(self._fixed_feature_dims.items()), 0,
'Expected no fixed features')
check.Eq(len(self._linked_feature_dims.items()), 2,
'Expected two linked features')
check.In('sources', self._linked_feature_dims,
'Missing required linked feature')
check.In('targets', self._linked_feature_dims,
'Missing required linked feature')
self._source_dim = self._linked_feature_dims['sources']
self._target_dim = self._linked_feature_dims['targets']
# TODO(googleuser): Make parameter initialization configurable.
self._weights = []
self._weights.append(
tf.get_variable(
'weights_pair',
[self._num_labels, self._source_dim, self._target_dim],
tf.float32,
tf.random_normal_initializer(stddev=1e-4, seed=self._seed)))
self._weights.append(
tf.get_variable(
'weights_source', [self._num_labels, self._source_dim],
tf.float32,
tf.random_normal_initializer(stddev=1e-4, seed=self._seed)))
self._weights.append(
tf.get_variable(
'weights_target', [self._num_labels, self._target_dim],
tf.float32,
tf.random_normal_initializer(stddev=1e-4, seed=self._seed)))
self._biases = []
self._biases.append(
tf.get_variable(
'biases', [self._num_labels], tf.float32,
tf.random_normal_initializer(stddev=1e-4, seed=self._seed)))
self._params.extend(self._weights + self._biases)
self._regularized_weights.extend(self._weights)
self._layers.append(
network_units.Layer(self, 'labels', self._num_labels))
def __init__(self, component):
"""Initializes weights and layers.
Args:
component: Parent ComponentBuilderBase object.
"""
super(BiaffineDigraphNetwork, self).__init__(component)
check.Eq(len(self._fixed_feature_dims.items()), 0,
'Expected no fixed features')
check.Eq(len(self._linked_feature_dims.items()), 2,
'Expected two linked features')
check.In('sources', self._linked_feature_dims,
'Missing required linked feature')
check.In('targets', self._linked_feature_dims,
'Missing required linked feature')
self._source_dim = self._linked_feature_dims['sources']
self._target_dim = self._linked_feature_dims['targets']
# TODO(googleuser): Make parameter initialization configurable.
self._weights = []
self._weights.append(
tf.get_variable(
'weights_arc', [self._source_dim, self._target_dim],
tf.float32,
tf.random_normal_initializer(stddev=1e-4, seed=self._seed)))
self._weights.append(
tf.get_variable(
'weights_source', [self._source_dim], tf.float32,
tf.random_normal_initializer(stddev=1e-4, seed=self._seed)))
self._weights.append(
tf.get_variable(
'root', [self._source_dim], tf.float32,
tf.random_normal_initializer(stddev=1e-4, seed=self._seed)))
self._params.extend(self._weights)
self._regularized_weights.extend(self._weights)
# Negative Layer.dim indicates that the dimension is dynamic.
self._layers.append(network_units.Layer(self, 'adjacency', -1))
def __init__(self, component):
"""Initializes weights and layers.
Args:
component: Parent ComponentBuilderBase object.
"""
super(BiaffineDigraphNetwork, self).__init__(component)
check.Eq(len(self._fixed_feature_dims.items()), 0,
'Expected no fixed features')
check.Eq(len(self._linked_feature_dims.items()), 2,
'Expected two linked features')
check.In('sources', self._linked_feature_dims,
'Missing required linked feature')
check.In('targets', self._linked_feature_dims,
'Missing required linked feature')
self._source_dim = self._linked_feature_dims['sources']
self._target_dim = self._linked_feature_dims['targets']
self._weights = []
self._weights.append(
tf.get_variable('weights_arc',
[self._source_dim, self._target_dim], tf.float32,
tf.orthogonal_initializer()))
self._weights.append(
tf.get_variable('weights_source', [self._source_dim], tf.float32,
tf.zeros_initializer()))
self._weights.append(
tf.get_variable('root', [self._source_dim], tf.float32,
tf.zeros_initializer()))
self._params.extend(self._weights)
self._regularized_weights.extend(self._weights)
# Add runtime hooks for pre-computed weights.
self._derived_params.append(self._get_root_weights)
self._derived_params.append(self._get_root_bias)
# Negative Layer.dim indicates that the dimension is dynamic.
self._layers.append(network_units.Layer(component, 'adjacency', -1))
def __init__(self, component):
super(BulkBiLSTMNetwork, self).__init__(component)
check.In('lengths', self._linked_feature_dims,
'Missing required linked feature')
check.Eq(self._linked_feature_dims['lengths'], 1,
'Wrong dimension for "lengths" feature')
self._input_dim = self._concatenated_input_dim - 1 # exclude 'lengths'
self._output_dim = self.get_layer_size('outputs')
tf.logging.info('[%s] Bulk bi-LSTM with input_dim=%d output_dim=%d',
component.name, self._input_dim, self._output_dim)
# Create one training and inference cell per layer and direction.
self._train_cells_forward = self._create_train_cells()
self._train_cells_backward = self._create_train_cells()
self._inference_cells_forward = self._create_inference_cells()
self._inference_cells_backward = self._create_inference_cells()
def _bilstm_closure(scope):
"""Applies the bi-LSTM to placeholder inputs and lengths."""
# Use singleton |stride| and |steps| because their values don't affect the
# weight variables.
stride, steps = 1, 1
placeholder_inputs = tf.placeholder(
dtype=tf.float32, shape=[stride, steps, self._input_dim])
placeholder_lengths = tf.placeholder(dtype=tf.int64,
shape=[stride])
# Omit the initial states and sequence lengths for simplicity; they don't
# affect the weight variables.
tf.contrib.rnn.stack_bidirectional_dynamic_rnn(
self._train_cells_forward,
self._train_cells_backward,
placeholder_inputs,
dtype=tf.float32,
sequence_length=placeholder_lengths,
scope=scope)
self._capture_variables_as_params(_bilstm_closure)
# Allocate parameters for the initial states. Note that an LSTM state is a
# tuple of two substates (c, h), so there are 4 variables per layer.
for index, num_units in enumerate(self._hidden_layer_sizes):
for direction in ['forward', 'backward']:
for substate in ['c', 'h']:
self._params.append(
tf.get_variable(
'initial_state_%s_%s_%d' %
(direction, substate, index),
[1, num_units
], # leading 1 for later batch-wise tiling
dtype=tf.float32,
initializer=tf.constant_initializer(0.0)))