def __init__(self, config, **kwargs):
super().__init__(config, **kwargs)
self.config = config
self.embedding_layer = EmbeddingSharedWeights.get_from_graph()
if config.editor.word_dim != config.editor.edit_encoder.extractor.hidden_size:
self.embedding_layer = self.embedding_layer.get_projected(
config.editor.edit_encoder.extractor.hidden_size)
self.micro_ev_projection = tf.layers.Dense(
config.editor.edit_encoder.micro_ev_dim,
activation=config.editor.edit_encoder.get('mev_proj_activation_fn',
None),
use_bias=True,
name='micro_ev_proj')
self.edit_vector_projection = tf.layers.Dense(
config.editor.edit_encoder.edit_dim,
activation=config.editor.edit_encoder.get(
'edit_vector_proj_activation_fn', None),
use_bias=False,
name='encoder_ev')
extractor_config = Config.merge_to_new(
[config.editor.transformer, config.editor.edit_encoder.extractor])
extractor_config.put('save_attentions',
config.get('eval.save_attentions', False))
self.mev_extractor = TransformerMicroEditExtractor(
self.embedding_layer, self.micro_ev_projection, extractor_config)
def __init__(self, config, **kwargs):
super().__init__(**kwargs)
self.config = Config.merge_to_new(
[config.editor.transformer, config.editor.decoder])
# change hidden_size of transformer to match with the augmented input
self.config.put('orig_hidden_size', self.config.hidden_size)
self.config.put(
'hidden_size',
self.config.hidden_size + config.editor.edit_encoder.edit_dim)
# Project embedding to transformer's hidden_size if needed
embedding_layer = EmbeddingSharedWeights.get_from_graph()
self.vocab_size = embedding_layer.vocab_size
if config.editor.word_dim != self.config.orig_hidden_size:
self.embedding_layer = embedding_layer.get_projected(
self.config.orig_hidden_size)
else:
self.embedding_layer = embedding_layer
# As far as EmbeddingSharedWeights class supports linear projection on embeddings,
# we will use it to compute the model's logits
self.project_back = tf.layers.Dense(config.editor.word_dim,
activation=None,
name='project_back')
self.vocab_projection = embedding_layer
# Transformer stack
self.decoder_stack = MultiSourceDecoderStack(self.config.to_json(),
graph_utils.is_training())
def __init__(self, config: Config):
self.config = config
encoder_config = Config.merge_to_new(
[config.editor.transformer, config.editor.encoder])
self.encoder = TransformerEncoder(encoder_config, name='encoder')
self.edit_encoder = EditEncoder(config)
self.decoder = Decoder(config)
def __init__(self, config, sent_encoder, **kwargs):
super().__init__(config, **kwargs)
extractor_config = Config.merge_to_new(
[config.editor.transformer, config.editor.edit_encoder.extractor])
extractor_config.put('save_attentions',
config.get('eval.save_attentions', False))
self.mev_extractor = TransformerMicroEditExtractor(
self.embedding_layer, self.micro_ev_projection, sent_encoder,
extractor_config)
def __init__(self, config, data_dir, checkpoint=None):
self._data_dir = Path(data_dir)
self._checkpoint = checkpoint
self._config = Config.from_file(config)
self._path_model_dir()
self._put_epoch_num()
self._config.put('local_data_dir', self._data_dir)
tf.logging.info("Model: " + self.model.NAME)
tf.logging.info("Config:")
tf.logging.info(str(self._config))
print()
def __init__(self, embedding_layer, mev_projection, params, **kwargs):
super().__init__(**kwargs)
self.params = params
is_training = graph_utils.is_training()
encoder_config = Config.merge_to_new([params, params.encoder])
decoder_config = Config.merge_to_new([params, params.decoder])
self.target_encoder = EncoderStack(encoder_config.to_json(),
is_training, params.save_attentions)
self.mev_decoder = DecoderStack(decoder_config.to_json(), is_training,
params.save_attentions)
self.embedding_layer = embedding_layer
self.mev_projection = mev_projection
self.cls_tok_embedding = self.add_weight('cls_tok_embedding',
(self.params.hidden_size, ),
dtype=tf.float32,
trainable=True)
self.pooling_layer = tf.layers.Dense(self.params.hidden_size,
activation='tanh',
name='pooling_layer')