def testSequenceToSequenceWithScheduledSampling(self):
model = models.SequenceToSequence(
inputters.WordEmbedder(16),
inputters.WordEmbedder(16),
encoders.SelfAttentionEncoder(2, 16, 4, 32),
decoders.RNNDecoder(2, 16))
params = {
"scheduled_sampling_type": "linear",
"scheduled_sampling_read_probability": 0.8,
"scheduled_sampling_k": 0.1
}
features_file, labels_file, data_config = self._makeToyEnDeData()
model.initialize(data_config, params=params)
dataset = model.examples_inputter.make_training_dataset(features_file, labels_file, 16)
features, labels = next(iter(dataset))
with self.assertRaises(ValueError):
model(features, labels=labels, training=True) # step argument is required.
outputs, _ = model(features, labels=labels, training=True, step=10)
self.assertEqual(outputs["logits"].shape[1], labels["ids"].shape[1])
def testRNNDecoder(self):
decoder = decoders.RNNDecoder(2, 20)
self._testDecoder(decoder, support_alignment_history=False)
def testRNNDecoder(self): decoder = decoders.RNNDecoder(2, 20) self._testDecoder(decoder)
def testRNNDecoderTraining(self):
decoder = decoders.RNNDecoder(2, 20)
self._testDecoderTraining(decoder)
