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)