def testBLSTM(self):
lstm = rnn.LSTM(3, 12, dropout=0.5, bidirectional=True)
inputs = tf.random.uniform([4, 5, 5])
outputs, states = lstm(inputs, training=True)
self.assertListEqual(outputs.shape.as_list(), [4, 5, 24])
self.assertIsInstance(states, tuple)
self.assertEqual(len(states), 3)
def testLSTM(self):
lstm = rnn.LSTM(3, 12)
inputs = tf.random.uniform([4, 5, 5])
outputs, states = lstm(inputs, training=True)
self.assertListEqual(outputs.shape.as_list(), [4, 5, 12])
self.assertIsInstance(states, tuple)
self.assertEqual(len(states), 3)
def __init__(self,
num_layers,
num_units,
bidirectional=False,
residual_connections=False,
dropout=0.3,
reducer=ConcatReducer(),
**kwargs):
"""Initializes the parameters of the encoder.
Args:
num_layers: The number of layers.
num_units: The number of units in each layer output.
bidirectional: Make each LSTM layer bidirectional.
residual_connections: If ``True``, each layer input will be added to its
output.
dropout: The probability to drop units in each layer output.
reducer: A :class:`opennmt.layers.Reducer` instance to merge
bidirectional state and outputs.
**kwargs: Additional layer arguments.
"""
lstm_layer = rnn.LSTM(
num_layers,
num_units,
bidirectional=bidirectional,
reducer=reducer,
dropout=dropout,
residual_connections=residual_connections,
)
super().__init__(lstm_layer, **kwargs)
def testLSTMWithMask(self):
lstm = rnn.LSTM(3, 12)
inputs = tf.random.uniform([3, 4, 5])
lengths = [4, 2, 3]
mask = tf.sequence_mask(lengths)
outputs, states = lstm(inputs, mask=mask, training=True)
last_hidden = states[-1][0]
for i, length in enumerate(lengths):
self.assertAllClose(last_hidden[i], outputs[i, length - 1])