def cross_attention(self, table_encodes, document_encodes, num_units, num_heads, num_layers, ffn_inner_dim, sequence_length=None, mode=tf.estimator.ModeKeys.TRAIN): table_encodes *= num_units**0.5 # if self.position_encoder is not None: # inputs = self.position_encoder(inputs) inputs = tf.layers.dropout( table_encodes, rate=FLAGS.attention_dropout, training=mode == tf.estimator.ModeKeys.TRAIN) mask = transformer.build_sequence_mask( sequence_length, num_heads=num_heads, maximum_length=tf.shape(document_encodes)[1]) state = () for l in range(num_layers): with tf.variable_scope("layer_{}".format(l)): with tf.variable_scope("multi_head"): context = transformer.multi_head_attention( num_heads, transformer.norm(inputs), document_encodes, mode, num_units=num_units, mask=mask, dropout=FLAGS.attention_dropout) context = transformer.drop_and_add(inputs, context, mode, dropout=FLAGS.dropout) with tf.variable_scope("ffn"): transformed = transformer.feed_forward( transformer.norm(context), ffn_inner_dim, mode, dropout=FLAGS.attention_dropout) transformed = transformer.drop_and_add( context, transformed, mode, dropout=FLAGS.dropout) inputs = transformed state += (tf.reduce_mean(inputs, axis=1), ) outputs = transformer.norm(inputs) # return (outputs, state, sequence_length) return outputs
def encode(self, inputs, sequence_length=None, mode=tf.estimator.ModeKeys.TRAIN): inputs *= self.num_units**0.5 if self.position_encoder is not None: inputs = self.position_encoder(inputs) else: print("===============================================") print("no position encoder") inputs = tf.layers.dropout( inputs, rate=self.dropout, training=mode == tf.estimator.ModeKeys.TRAIN) mask = transformer.build_sequence_mask( sequence_length, num_heads=self.num_heads, maximum_length=tf.shape(inputs)[1]) state = () for l in range(self.num_layers): with tf.variable_scope("layer_{}".format(l)): with tf.variable_scope("multi_head"): context = transformer.multi_head_attention( self.num_heads, transformer.norm(inputs), None, mode, num_units=self.num_units, mask=mask, dropout=self.attention_dropout) context = transformer.drop_and_add(inputs, context, mode, dropout=self.dropout) with tf.variable_scope("ffn"): transformed = transformer.feed_forward( transformer.norm(context), self.ffn_inner_dim, mode, dropout=self.relu_dropout) transformed = transformer.drop_and_add( context, transformed, mode, dropout=self.dropout) inputs = transformed state += (tf.reduce_mean(inputs, axis=1), ) outputs = transformer.norm(inputs) return (outputs, state, sequence_length)
def encode(self, inputs, sequence_length=None, mode=tf.estimator.ModeKeys.TRAIN): inputs *= self.num_units**0.5 if self.position_encoder is not None: inputs = self.position_encoder(inputs, sequence_length=sequence_length) inputs = tf.layers.dropout( inputs, rate=self.dropout, training=mode == tf.estimator.ModeKeys.TRAIN) mask = transformer.build_sequence_mask( sequence_length, num_heads=self.num_heads, maximum_length=tf.shape(inputs)[1], dtype=inputs.dtype) state = () for l in range(self.num_layers): with tf.variable_scope("layer_{}".format(l)): with tf.variable_scope("multi_head"): context = transformer.multi_head_attention( self.num_heads, transformer.norm(inputs), None, mode, num_units=self.num_units, mask=mask, dropout=self.attention_dropout) context = transformer.drop_and_add( inputs, context, mode, dropout=self.dropout) with tf.variable_scope("ffn"): transformed = transformer.feed_forward( transformer.norm(context), self.ffn_inner_dim, mode, dropout=self.relu_dropout) transformed = transformer.drop_and_add( context, transformed, mode, dropout=self.dropout) inputs = transformed state += (tf.reduce_mean(inputs, axis=1),) outputs = transformer.norm(inputs) return (outputs, state, sequence_length)
def _self_attention_stack(self, inputs, sequence_length=None, mode=tf.estimator.ModeKeys.TRAIN, cache=None, memory=None, memory_sequence_length=None): inputs = tf.layers.dropout( inputs, rate=self.dropout, training=mode == tf.estimator.ModeKeys.TRAIN) decoder_mask = None memory_mask = None if sequence_length is not None: decoder_mask = transformer.build_future_mask( sequence_length, num_heads=self.num_heads, maximum_length=tf.shape(inputs)[1], dtype=inputs.dtype) if memory is not None: if cache is not None: memory_mask = cache["memory_mask"] elif memory_sequence_length is not None: memory_mask = self._build_memory_mask( memory, memory_sequence_length=memory_sequence_length) for l in range(self.num_layers): layer_name = "layer_{}".format(l) layer_cache = cache[layer_name] if cache is not None else None with tf.variable_scope(layer_name): with tf.variable_scope("masked_multi_head"): encoded = transformer.multi_head_attention( self.num_heads, transformer.norm(inputs), None, mode, num_units=self.num_units, mask=decoder_mask, cache=layer_cache, dropout=self.attention_dropout) encoded = transformer.drop_and_add(inputs, encoded, mode, dropout=self.dropout) if memory is not None: with tf.variable_scope("multi_head"): context = transformer.multi_head_attention( self.num_heads, transformer.norm(encoded), memory, mode, mask=memory_mask, cache=layer_cache, dropout=self.attention_dropout) context = transformer.drop_and_add( encoded, context, mode, dropout=self.dropout) with tf.variable_scope("ffn"): transformed = transformer.feed_forward( transformer.norm(context), self.ffn_inner_dim, mode, dropout=self.relu_dropout) transformed = transformer.drop_and_add( context, transformed, mode, dropout=self.dropout) inputs = transformed outputs = transformer.norm(inputs) return outputs
def _self_attention_stack(self, inputs, sequence_length=None, mode=tf.estimator.ModeKeys.TRAIN, cache=None, memory=None, memory_sequence_length=None, step=None): inputs *= self.num_units**0.5 if self.position_encoder is not None: if step is None: inputs = self.position_encoder(inputs, sequence_length=sequence_length) else: inputs = self.position_encoder.apply_one(inputs, step + 1) inputs = tf.layers.dropout( inputs, rate=self.dropout, training=mode == tf.estimator.ModeKeys.TRAIN) decoder_mask = None memory_mask = None last_attention = None if self.self_attention_type == "scaled_dot": if sequence_length is not None: decoder_mask = transformer.build_future_mask( sequence_length, num_heads=self.num_heads, maximum_length=tf.shape(inputs)[1]) elif self.self_attention_type == "average": if cache is None: if sequence_length is None: sequence_length = tf.fill([tf.shape(inputs)[0]], tf.shape(inputs)[1]) decoder_mask = transformer.cumulative_average_mask( sequence_length, maximum_length=tf.shape(inputs)[1], dtype=inputs.dtype) if memory is not None and memory_sequence_length is not None: memory_mask = transformer.build_sequence_mask( memory_sequence_length, num_heads=self.num_heads, maximum_length=tf.shape(memory)[1]) for l in range(self.num_layers): layer_name = "layer_{}".format(l) layer_cache = cache[layer_name] if cache is not None else None with tf.variable_scope(layer_name): if self.self_attention_type == "scaled_dot": with tf.variable_scope("masked_multi_head"): encoded = transformer.multi_head_attention( self.num_heads, transformer.norm(inputs), None, mode, num_units=self.num_units, mask=decoder_mask, cache=layer_cache, dropout=self.attention_dropout) encoded = transformer.drop_and_add( inputs, encoded, mode, dropout=self.dropout) elif self.self_attention_type == "average": with tf.variable_scope("average_attention"): # Cumulative average. x = transformer.norm(inputs) y = transformer.cumulative_average( x, decoder_mask if cache is None else step, cache=layer_cache) # FFN. y = transformer.feed_forward( y, self.ffn_inner_dim, mode, dropout=self.relu_dropout) # Gating layer. z = tf.layers.dense(tf.concat([x, y], -1), self.num_units * 2) i, f = tf.split(z, 2, axis=-1) y = tf.sigmoid(i) * x + tf.sigmoid(f) * y encoded = transformer.drop_and_add( inputs, y, mode, dropout=self.dropout) if memory is not None: with tf.variable_scope("multi_head"): context, last_attention = transformer.multi_head_attention( self.num_heads, transformer.norm(encoded), memory, mode, mask=memory_mask, cache=layer_cache, dropout=self.attention_dropout, return_attention=True) context = transformer.drop_and_add( encoded, context, mode, dropout=self.dropout) else: context = encoded with tf.variable_scope("ffn"): transformed = transformer.feed_forward( transformer.norm(context), self.ffn_inner_dim, mode, dropout=self.relu_dropout) transformed = transformer.drop_and_add( context, transformed, mode, dropout=self.dropout) inputs = transformed if last_attention is not None: # The first head of the last layer is returned. first_head_attention = last_attention[:, 0] else: first_head_attention = None outputs = transformer.norm(inputs) return outputs, first_head_attention
def _self_attention_stack(self, inputs, sequence_length=None, mode=tf.estimator.ModeKeys.TRAIN, cache=None, memory=None, memory_sequence_length=None): inputs = tf.layers.dropout( inputs, rate=self.dropout, training=mode == tf.estimator.ModeKeys.TRAIN) decoder_mask = None memory_mask = None if sequence_length is not None: decoder_mask = transformer.build_future_mask( sequence_length, num_heads=self.num_heads, maximum_length=tf.shape(inputs)[1], dtype=inputs.dtype) if memory is not None: if cache is not None: memory_mask = cache["memory_mask"] elif memory_sequence_length is not None: memory_mask = self._build_memory_mask( memory, memory_sequence_length=memory_sequence_length) for l in range(self.num_layers): layer_name = "layer_{}".format(l) layer_cache = cache[layer_name] if cache is not None else None with tf.variable_scope(layer_name): with tf.variable_scope("masked_multi_head"): encoded = transformer.multi_head_attention( self.num_heads, transformer.norm(inputs), None, mode, num_units=self.num_units, mask=decoder_mask, cache=layer_cache, dropout=self.attention_dropout) encoded = transformer.drop_and_add( inputs, encoded, mode, dropout=self.dropout) if memory is not None: with tf.variable_scope("multi_head"): context = transformer.multi_head_attention( self.num_heads, transformer.norm(encoded), memory, mode, mask=memory_mask, cache=layer_cache, dropout=self.attention_dropout) context = transformer.drop_and_add( encoded, context, mode, dropout=self.dropout) with tf.variable_scope("ffn"): transformed = transformer.feed_forward( transformer.norm(context), self.ffn_inner_dim, mode, dropout=self.relu_dropout) transformed = transformer.drop_and_add( context, transformed, mode, dropout=self.dropout) inputs = transformed outputs = transformer.norm(inputs) return outputs
def _self_attention_stack(self, inputs, # batch, max_dec_len, emb_dim sequence_length=None, # [batch] mode=tf.estimator.ModeKeys.TRAIN, cache=None, memory=None, # [batch, enc_len, num_units] memory_sequence_length=None, # [batch] step=None): inputs *= self.num_units ** 0.5 if self.position_encoder is not None: inputs = self.position_encoder(inputs, position=step + 1 if step is not None else None) # inputs [batch, max_dec_len, emb_dim] inputs = tf.layers.dropout( # batch, max_dec_len, emb_dim inputs, rate=self.dropout, training=mode == tf.estimator.ModeKeys.TRAIN) decoder_mask = None memory_mask = None last_attention = None if self.self_attention_type == "scaled_dot": if sequence_length is not None: # sequence_length is None when decode, not None at train decoder_mask = transformer.build_future_mask( # [batch, 1, max_dec_len, max_dec_len] sequence_length, num_heads=self.num_heads, maximum_length=tf.shape(inputs)[1]) elif self.self_attention_type == "average": if cache is None: if sequence_length is None: sequence_length = tf.fill([tf.shape(inputs)[0]], tf.shape(inputs)[1]) decoder_mask = transformer.cumulative_average_mask( sequence_length, maximum_length=tf.shape(inputs)[1], dtype=inputs.dtype) if memory is not None and not tf.contrib.framework.nest.is_sequence(memory): memory = (memory,) if memory_sequence_length is not None: if not tf.contrib.framework.nest.is_sequence(memory_sequence_length): memory_sequence_length = (memory_sequence_length,) memory_mask = [ # [batch, 1, 1, enc_len] transformer.build_sequence_mask( length, num_heads=self.num_heads, maximum_length=tf.shape(m)[1]) for m, length in zip(memory, memory_sequence_length)] for l in range(self.num_layers): layer_name = "layer_{}".format(l) layer_cache = cache[layer_name] if cache is not None else None # train的时候没有cache,decode的时候有cache with tf.variable_scope(layer_name): # self attention encode the decoder input (training) or last step output (decode) if self.self_attention_type == "scaled_dot": with tf.variable_scope("masked_multi_head"): encoded = transformer.multi_head_attention( # [batch, decode_len, hidden] self.num_heads, transformer.norm(inputs), None, mode, num_units=self.num_units, mask=decoder_mask, # [batch, 1, len, len] cache=layer_cache, dropout=self.attention_dropout) last_context = transformer.drop_and_add( # [batch, decode_len, hidden] inputs, encoded, mode, dropout=self.dropout) elif self.self_attention_type == "average": with tf.variable_scope("average_attention"): # Cumulative average. x = transformer.norm(inputs) y = transformer.cumulative_average( x, decoder_mask if cache is None else step, cache=layer_cache) # FFN. y = transformer.feed_forward( y, self.ffn_inner_dim, mode, dropout=self.relu_dropout) # Gating layer. z = tf.layers.dense(tf.concat([x, y], -1), self.num_units * 2) i, f = tf.split(z, 2, axis=-1) y = tf.sigmoid(i) * x + tf.sigmoid(f) * y last_context = transformer.drop_and_add( inputs, y, mode, dropout=self.dropout) # attending to encoder memory using decoder context if memory is not None: for i, (mem, mask) in enumerate(zip(memory, memory_mask)): memory_cache = layer_cache["memory"][ i] if layer_cache is not None else None # train的时候没有cache,decode的时候有cache with tf.variable_scope("multi_head" if i == 0 else "multi_head_%d" % i): context, last_attention = transformer.multi_head_attention( self.num_heads, transformer.norm(last_context), mem, # [batch, enc_len, dim] mode, mask=mask, # [batch, 1, 1, len] cache=memory_cache, dropout=self.attention_dropout, return_attention=True) # context [batch, decode_len, num_units], last_attention train[batch, head, dec_len, enc_len] last_context = transformer.drop_and_add( last_context, # [batch, decode_len, num_units] context, mode, dropout=self.dropout) if i > 0: # Do not return attention in case of multi source. last_attention = None with tf.variable_scope("ffn"): transformed = transformer.feed_forward( # [batch, decode_len, num_units] transformer.norm(last_context), self.ffn_inner_dim, mode, dropout=self.relu_dropout) transformed = transformer.drop_and_add( # [batch, decode_len, num_units] last_context, transformed, mode, dropout=self.dropout) inputs = transformed if last_attention is not None: # The first head of the last layer is returned. first_head_attention = last_attention[:, 0] else: first_head_attention = None outputs = transformer.norm(inputs) # [batch, decode_len, num_units] return outputs, first_head_attention
def encode(self, inputs, sequence_length=None, mode=tf.estimator.ModeKeys.TRAIN): """ :param inputs: [batch, enc_len, emb_dim] :param sequence_length: [batch] :param mode: :return: outputs: [batch, len, dim] last layer output state: a tuple ([batch, dim]) * num_layers, contains the sum over len of each layer outputs sequence_length [batch] """ inputs *= self.num_units**0.5 if self.position_encoder is not None: inputs = self.position_encoder(inputs) inputs = tf.layers.dropout( inputs, rate=self.dropout, training=mode == tf.estimator.ModeKeys.TRAIN) mask = transformer.build_sequence_mask( # [batch, 1, 1, enc_len] sequence_length, num_heads=self.num_heads, maximum_length=tf.shape(inputs)[1]) state = () for l in range(self.num_layers): with tf.variable_scope("layer_{}".format(l)): with tf.variable_scope("multi_head"): context = transformer.multi_head_attention( # [batch, len, dim] self.num_heads, transformer.norm(inputs), None, mode, num_units=self.num_units, mask=mask, dropout=self.attention_dropout) context = transformer.drop_and_add( # [batch, len, dim] inputs, context, mode, dropout=self.dropout) with tf.variable_scope("ffn"): transformed = transformer.feed_forward( # [batch, len, dim] transformer.norm(context), self.ffn_inner_dim, mode, dropout=self.relu_dropout) transformed = transformer.drop_and_add( # [batch, len, dim] context, transformed, mode, dropout=self.dropout) inputs = transformed state += (tf.reduce_mean(inputs, axis=1), ) outputs = transformer.norm(inputs) # [batch, len, dim] return (outputs, state, sequence_length)