def _mlp(self) -> MultilayerPerceptron:
mlp_input = tf.concat([enc.output for enc in self.encoders], 1)
return MultilayerPerceptron(mlp_input,
self.layers,
self.dropout_keep_prob,
len(self.vocabulary),
activation_fn=self.activation_fn,
train_mode=self.train_mode)
def __init__(self,
name: str,
encoders: List[Stateful],
vocabulary: Vocabulary,
data_id: str,
layers: List[int],
activation_fn: Callable[[tf.Tensor], tf.Tensor] = tf.nn.relu,
dropout_keep_prob: float = 0.5,
reuse: ModelPart = None,
save_checkpoint: str = None,
load_checkpoint: str = None,
initializers: InitializerSpecs = None) -> None:
"""Construct a new instance of the sequence classifier.
Args:
name: Name of the decoder. Should be unique accross all Neural
Monkey objects
encoders: Input encoders of the decoder
vocabulary: Target vocabulary
data_id: Target data series
layers: List defining structure of the NN. Ini example:
layers=[100,20,5] ;creates classifier with hidden layers of
size 100, 20, 5 and one output layer
depending on the size of vocabulary
activation_fn: activation function used on the output of each
hidden layer.
dropout_keep_prob: Probability of keeping a value during dropout
"""
check_argument_types()
ModelPart.__init__(self, name, reuse, save_checkpoint, load_checkpoint,
initializers)
self.encoders = encoders
self.vocabulary = vocabulary
self.data_id = data_id
self.layers = layers
self.activation_fn = activation_fn
self.dropout_keep_prob = dropout_keep_prob
self.max_output_len = 1
with self.use_scope():
self.gt_inputs = [tf.placeholder(tf.int32, [None], "targets")]
mlp_input = tf.concat([enc.output for enc in self.encoders], 1)
self._mlp = MultilayerPerceptron(mlp_input,
self.layers,
self.dropout_keep_prob,
len(self.vocabulary),
activation_fn=self.activation_fn,
train_mode=self.train_mode)
tf.summary.scalar("train_optimization_cost",
self.cost,
collections=["summary_train"])
def __init__(self,
name: str,
encoders: List[Any],
vocabulary: Vocabulary,
data_id: str,
layers: Optional[List[int]] = None,
activation: Callable[[tf.Tensor], tf.Tensor] = tf.tanh,
dropout_keep_prob: float = 0.5,
save_checkpoint: Optional[str] = None,
load_checkpoint: Optional[str] = None) -> None:
ModelPart.__init__(self, name, save_checkpoint, load_checkpoint)
self.encoders = encoders
self.vocabulary = vocabulary
self.data_id = data_id
self.layers = layers
self.activation = activation
self.dropout_keep_prob = dropout_keep_prob
self.max_output_len = 1
with tf.variable_scope(name):
self.learning_step = tf.get_variable(
"learning_step", [],
trainable=False,
initializer=tf.constant_initializer(0))
self.dropout_placeholder = \
tf.placeholder(tf.float32, name="dropout_plc")
self.gt_inputs = [
tf.placeholder(tf.int32, shape=[None], name="targets")
]
mlp_input = tf.concat(1, [enc.encoded for enc in encoders])
mlp = MultilayerPerceptron(mlp_input,
layers, self.dropout_placeholder,
len(vocabulary))
self.loss_with_gt_ins = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
mlp.logits, self.gt_inputs[0]))
self.loss_with_decoded_ins = self.loss_with_gt_ins
self.cost = self.loss_with_gt_ins
self.decoded_seq = [mlp.classification]
self.decoded_logits = [mlp.logits]
self.runtime_logprobs = [tf.nn.log_softmax(mlp.logits)]
tf.scalar_summary('val_optimization_cost',
self.cost,
collections=["summary_val"])
tf.scalar_summary('train_optimization_cost',
self.cost,
collections=["summary_train"])
def __init__(self,
encoders,
vocabulary,
data_id,
name,
layers=[],
activation=tf.tanh,
dropout_keep_p=0.5):
self.encoders = encoders
self.vocabulary = vocabulary
self.data_id = data_id
self.layers = layers
self.activation = activation
self.dropout_keep_p = dropout_keep_p
self.name = name
self.max_output_len = 1
with tf.variable_scope(name):
self.learning_step = tf.get_variable(
"learning_step", [],
trainable=False,
initializer=tf.constant_initializer(0))
self.dropout_placeholder = tf.placeholder(tf.float32,
name="dropout_plc")
self.gt_inputs = [
tf.placeholder(tf.int32, shape=[None], name="targets")
]
mlp_input = tf.concat(1, [enc.encoded for enc in encoders])
mlp = MultilayerPerceptron(mlp_input,
layers, self.dropout_placeholder,
len(vocabulary))
self.loss_with_gt_ins = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
mlp.logits, self.gt_inputs[0]))
self.loss_with_decoded_ins = self.loss_with_gt_ins
self.cost = self.loss_with_gt_ins
self.decoded_seq = [mlp.classification]
self.decoded_logits = [mlp.logits]
tf.scalar_summary('val_optimization_cost',
self.cost,
collections=["summary_val"])
tf.scalar_summary('train_optimization_cost',
self.cost,
collections=["summary_train"])
def __init__(self,
name: str,
encoders: List[Any],
vocabulary: Vocabulary,
data_id: str,
layers: List[int],
activation_fn: Callable[[tf.Tensor], tf.Tensor] = tf.nn.relu,
dropout_keep_prob: float = 0.5,
save_checkpoint: Optional[str] = None,
load_checkpoint: Optional[str] = None) -> None:
"""Construct a new instance of the sequence classifier.
Args:
name: Name of the decoder. Should be unique accross all Neural
Monkey objects
encoders: Input encoders of the decoder
vocabulary: Target vocabulary
data_id: Target data series
layers: List defining structure of the NN. Ini example:
layers=[100,20,5] ;creates classifier with hidden layers of
size 100, 20, 5 and one output layer
depending on the size of vocabulary
activation_fn: activation function used on the output of each
hidden layer.
dropout_keep_prob: Probability of keeping a value during dropout
"""
ModelPart.__init__(self, name, save_checkpoint, load_checkpoint)
self.encoders = encoders
self.vocabulary = vocabulary
self.data_id = data_id
self.layers = layers
self.activation_fn = activation_fn
self.dropout_keep_prob = dropout_keep_prob
self.max_output_len = 1
with self.use_scope():
self.learning_step = tf.get_variable(
"learning_step", [],
trainable=False,
initializer=tf.constant_initializer(0))
self.dropout_placeholder = \
tf.placeholder(tf.float32, name="dropout_plc")
self.gt_inputs = [
tf.placeholder(tf.int32, shape=[None], name="targets")
]
mlp_input = tf.concat([enc.encoded for enc in encoders], 1)
mlp = MultilayerPerceptron(mlp_input,
layers,
self.dropout_placeholder,
len(vocabulary),
activation_fn=self.activation_fn)
self.loss_with_gt_ins = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=mlp.logits, labels=self.gt_inputs[0]))
self.loss_with_decoded_ins = self.loss_with_gt_ins
self.cost = self.loss_with_gt_ins
self.decoded_seq = [mlp.classification]
self.decoded_logits = [mlp.logits]
self.runtime_logprobs = [tf.nn.log_softmax(mlp.logits)]
tf.summary.scalar('val_optimization_cost',
self.cost,
collections=["summary_val"])
tf.summary.scalar('train_optimization_cost',
self.cost,
collections=["summary_train"])