seq2seq_model.py

# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

"""Sequence-to-sequence model with an attention mechanism."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import random

import numpy as np
from six.moves import xrange  # pylint: disable=redefined-builtin
import tensorflow as tf

import data_utils
from tensorflow.contrib import rnn
from tensorflow.contrib.framework import nest
from tensorflow.contrib.legacy_seq2seq import embedding_attention_decoder

def embedding_attention_seq2seq(encoder_inputs,
                                decoder_inputs,
                                enc_cell,
                                dec_cell,
                                num_encoder_symbols,
                                num_decoder_symbols,
                                embedding_size,
                                num_heads=1,
                                output_projection=None,
                                feed_previous=False,
                                dtype=None,
                                scope=None,
                                initial_state_attention=False):
  """Embedding sequence-to-sequence model with attention.

  This model first embeds encoder_inputs by a newly created embedding (of shape
  [num_encoder_symbols x input_size]). Then it runs an RNN to encode
  embedded encoder_inputs into a state vector. It keeps the outputs of this
  RNN at every step to use for attention later. Next, it embeds decoder_inputs
  by another newly created embedding (of shape [num_decoder_symbols x
  input_size]). Then it runs attention decoder, initialized with the last
  encoder state, on embedded decoder_inputs and attending to encoder outputs.

  Warning: when output_projection is None, the size of the attention vectors
  and variables will be made proportional to num_decoder_symbols, can be large.

  Args:
    encoder_inputs: A list of 1D int32 Tensors of shape [batch_size].
    decoder_inputs: A list of 1D int32 Tensors of shape [batch_size].
    cell: tf.nn.rnn_cell.RNNCell defining the cell function and size.
    num_encoder_symbols: Integer; number of symbols on the encoder side.
    num_decoder_symbols: Integer; number of symbols on the decoder side.
    embedding_size: Integer, the length of the embedding vector for each symbol.
    num_heads: Number of attention heads that read from attention_states.
    output_projection: None or a pair (W, B) of output projection weights and
      biases; W has shape [output_size x num_decoder_symbols] and B has
      shape [num_decoder_symbols]; if provided and feed_previous=True, each
      fed previous output will first be multiplied by W and added B.
    feed_previous: Boolean or scalar Boolean Tensor; if True, only the first
      of decoder_inputs will be used (the "GO" symbol), and all other decoder
      inputs will be taken from previous outputs (as in embedding_rnn_decoder).
      If False, decoder_inputs are used as given (the standard decoder case).
    dtype: The dtype of the initial RNN state (default: tf.float32).
    scope: VariableScope for the created subgraph; defaults to
      "embedding_attention_seq2seq".
    initial_state_attention: If False (default), initial attentions are zero.
      If True, initialize the attentions from the initial state and attention
      states.

  Returns:
    A tuple of the form (outputs, state), where:
      outputs: A list of the same length as decoder_inputs of 2D Tensors with
        shape [batch_size x num_decoder_symbols] containing the generated
        outputs.
      state: The state of each decoder cell at the final time-step.
        It is a 2D Tensor of shape [batch_size x cell.state_size].
  """
  with tf.variable_scope(
      scope or "embedding_attention_seq2seq", dtype=dtype) as scope:
    dtype = scope.dtype
    # Encoder.

    encoder_cell = enc_cell

    encoder_cell = rnn.EmbeddingWrapper(
        encoder_cell,
        embedding_classes=num_encoder_symbols,
        embedding_size=embedding_size)
    encoder_outputs, encoder_state = rnn.static_rnn(
        encoder_cell, encoder_inputs, dtype=dtype)

    # First calculate a concatenation of encoder outputs to put attention on.
    top_states = [
        tf.reshape(e, [-1, 1, encoder_cell.output_size]) for e in encoder_outputs
    ]
    attention_states = tf.concat(top_states, 1)

    # Decoder.
    output_size = None
    if output_projection is None:
      dec_cell = rnn.OutputProjectionWrapper(dec_cell, num_decoder_symbols)
      output_size = num_decoder_symbols

    if isinstance(feed_previous, bool):
      return embedding_attention_decoder(
          decoder_inputs,
          encoder_state,
          attention_states,
          dec_cell,
          num_decoder_symbols,
          embedding_size,
          num_heads=num_heads,
          output_size=output_size,
          output_projection=output_projection,
          feed_previous=feed_previous,
          initial_state_attention=initial_state_attention)

    # If feed_previous is a Tensor, we construct 2 graphs and use cond.
    def decoder(feed_previous_bool):
      reuse = None if feed_previous_bool else True
      with tf.variable_scope(
          tf.get_variable_scope(), reuse=reuse):
        outputs, state = embedding_attention_decoder(
            decoder_inputs,
            encoder_state,
            attention_states,
            dec_cell,
            num_decoder_symbols,
            embedding_size,
            num_heads=num_heads,
            output_size=output_size,
            output_projection=output_projection,
            feed_previous=feed_previous_bool,
            update_embedding_for_previous=False,
            initial_state_attention=initial_state_attention)
        state_list = [state]
        if nest.is_sequence(state):
          state_list = nest.flatten(state)
        return outputs + state_list

    outputs_and_state = tf.cond(feed_previous,
                                              lambda: decoder(True),
                                              lambda: decoder(False))
    outputs_len = len(decoder_inputs)  # Outputs length same as decoder inputs.
    state_list = outputs_and_state[outputs_len:]
    state = state_list[0]
    if nest.is_sequence(encoder_state):
      state = nest.pack_sequence_as(
          structure=encoder_state, flat_sequence=state_list)
    return outputs_and_state[:outputs_len], state


class Seq2SeqModel(object):
  """Sequence-to-sequence model with attention and for multiple buckets.

  This class implements a multi-layer recurrent neural network as encoder,
  and an attention-based decoder. This is the same as the model described in
  this paper: http://arxiv.org/abs/1412.7449 - please look there for details,
  or into the seq2seq library for complete model implementation.
  This class also allows to use GRU cells in addition to LSTM cells, and
  sampled softmax to handle large output vocabulary size. A single-layer
  version of this model, but with bi-directional encoder, was presented in
    http://arxiv.org/abs/1409.0473
  and sampled softmax is described in Section 3 of the following paper.
    http://arxiv.org/abs/1412.2007
  """

  def __init__(self,
               source_vocab_size,
               target_vocab_size,
               buckets,
               size,
               num_layers,
               max_gradient_norm,
               batch_size,
               learning_rate,
               learning_rate_decay_factor,
               use_lstm=False,
               num_samples=512,
               forward_only=False,
               dtype=tf.float32):
    """Create the model.

    Args:
      source_vocab_size: size of the source vocabulary.
      target_vocab_size: size of the target vocabulary.
      buckets: a list of pairs (I, O), where I specifies maximum input length
        that will be processed in that bucket, and O specifies maximum output
        length. Training instances that have inputs longer than I or outputs
        longer than O will be pushed to the next bucket and padded accordingly.
        We assume that the list is sorted, e.g., [(2, 4), (8, 16)].
      size: number of units in each layer of the model.
      num_layers: number of layers in the model.
      max_gradient_norm: gradients will be clipped to maximally this norm.
      batch_size: the size of the batches used during training;
        the model construction is independent of batch_size, so it can be
        changed after initialization if this is convenient, e.g., for decoding.
      learning_rate: learning rate to start with.
      learning_rate_decay_factor: decay learning rate by this much when needed.
      use_lstm: if true, we use LSTM cells instead of GRU cells.
      num_samples: number of samples for sampled softmax.
      forward_only: if set, we do not construct the backward pass in the model.
      dtype: the data type to use to store internal variables.
    """
    self.source_vocab_size = source_vocab_size
    self.target_vocab_size = target_vocab_size
    self.buckets = buckets
    self.batch_size = batch_size
    self.learning_rate = tf.Variable(
        float(learning_rate), trainable=False, dtype=dtype)
    self.learning_rate_decay_op = self.learning_rate.assign(
        self.learning_rate * learning_rate_decay_factor)
    self.global_step = tf.Variable(0, trainable=False)

    # If we use sampled softmax, we need an output projection.
    output_projection = None
    softmax_loss_function = None
    # Sampled softmax only makes sense if we sample less than vocabulary size.
    if num_samples > 0 and num_samples < self.target_vocab_size:
      w_t = tf.get_variable("proj_w", [self.target_vocab_size, size], dtype=dtype)
      w = tf.transpose(w_t)
      b = tf.get_variable("proj_b", [self.target_vocab_size], dtype=dtype)
      output_projection = (w, b)

      def sampled_loss(labels, logits):
        labels = tf.reshape(labels, [-1, 1])
        # We need to compute the sampled_softmax_loss using 32bit floats to
        # avoid numerical instabilities.
        local_w_t = tf.cast(w_t, tf.float32)
        local_b = tf.cast(b, tf.float32)
        local_inputs = tf.cast(logits, tf.float32)
        return tf.cast(
            tf.nn.sampled_softmax_loss(
                weights=local_w_t,
                biases=local_b,
                labels=labels,
                inputs=local_inputs,
                num_sampled=num_samples,
                num_classes=self.target_vocab_size),
            dtype)
      softmax_loss_function = sampled_loss

    # Create the internal multi-layer cell for our RNN.
    def single_cell():
      return tf.contrib.rnn.GRUCell(size)
    if use_lstm:
      def single_cell():
        return tf.contrib.rnn.BasicLSTMCell(size)
    cell = single_cell()
    if num_layers > 1:
      cell = tf.contrib.rnn.MultiRNNCell([single_cell() for _ in range(num_layers)])

    # The seq2seq function: we use embedding for the input and attention.
    def seq2seq_f(encoder_inputs, decoder_inputs, do_decode):
      return embedding_attention_seq2seq(
          encoder_inputs,
          decoder_inputs,
          cell,
          cell,
          num_encoder_symbols=source_vocab_size,
          num_decoder_symbols=target_vocab_size,
          embedding_size=size,
          output_projection=output_projection,
          feed_previous=do_decode,
          dtype=dtype)

    # Feeds for inputs.
    self.encoder_inputs = []
    self.decoder_inputs = []
    self.target_weights = []
    for i in xrange(buckets[-1][0]):  # Last bucket is the biggest one.
      self.encoder_inputs.append(tf.placeholder(tf.int32, shape=[None],
                                                name="encoder{0}".format(i)))
    for i in xrange(buckets[-1][1] + 1):
      self.decoder_inputs.append(tf.placeholder(tf.int32, shape=[None],
                                                name="decoder{0}".format(i)))
      self.target_weights.append(tf.placeholder(dtype, shape=[None],
                                                name="weight{0}".format(i)))

    # Our targets are decoder inputs shifted by one.
    targets = [self.decoder_inputs[i + 1]
               for i in xrange(len(self.decoder_inputs) - 1)]

    # Training outputs and losses.
    if forward_only:
      self.outputs, self.losses = tf.contrib.legacy_seq2seq.model_with_buckets(
          self.encoder_inputs, self.decoder_inputs, targets,
          self.target_weights, buckets, lambda x, y: seq2seq_f(x, y, True),
          softmax_loss_function=softmax_loss_function)
      # If we use output projection, we need to project outputs for decoding.
      if output_projection is not None:
        for b in xrange(len(buckets)):
          self.outputs[b] = [
              tf.matmul(output, output_projection[0]) + output_projection[1]
              for output in self.outputs[b]
          ]
    else:
      self.outputs, self.losses = tf.contrib.legacy_seq2seq.model_with_buckets(
          self.encoder_inputs, self.decoder_inputs, targets,
          self.target_weights, buckets,
          lambda x, y: seq2seq_f(x, y, False),
          softmax_loss_function=softmax_loss_function)

    # Gradients and SGD update operation for training the model.
    params = tf.trainable_variables()
    if not forward_only:
      self.gradient_norms = []
      self.updates = []
      opt = tf.train.GradientDescentOptimizer(self.learning_rate)
      for b in xrange(len(buckets)):
        gradients = tf.gradients(self.losses[b], params)
        clipped_gradients, norm = tf.clip_by_global_norm(gradients,
                                                         max_gradient_norm)
        self.gradient_norms.append(norm)
        self.updates.append(opt.apply_gradients(
            zip(clipped_gradients, params), global_step=self.global_step))

    self.saver = tf.train.Saver(tf.global_variables())

  def step(self, session, encoder_inputs, decoder_inputs, target_weights,
           bucket_id, forward_only):
    """Run a step of the model feeding the given inputs.

    Args:
      session: tensorflow session to use.
      encoder_inputs: list of numpy int vectors to feed as encoder inputs.
      decoder_inputs: list of numpy int vectors to feed as decoder inputs.
      target_weights: list of numpy float vectors to feed as target weights.
      bucket_id: which bucket of the model to use.
      forward_only: whether to do the backward step or only forward.

    Returns:
      A triple consisting of gradient norm (or None if we did not do backward),
      average perplexity, and the outputs.

    Raises:
      ValueError: if length of encoder_inputs, decoder_inputs, or
        target_weights disagrees with bucket size for the specified bucket_id.
    """
    # Check if the sizes match.
    encoder_size, decoder_size = self.buckets[bucket_id]
    if len(encoder_inputs) != encoder_size:
      raise ValueError("Encoder length must be equal to the one in bucket,"
                       " %d != %d." % (len(encoder_inputs), encoder_size))
    if len(decoder_inputs) != decoder_size:
      raise ValueError("Decoder length must be equal to the one in bucket,"
                       " %d != %d." % (len(decoder_inputs), decoder_size))
    if len(target_weights) != decoder_size:
      raise ValueError("Weights length must be equal to the one in bucket,"
                       " %d != %d." % (len(target_weights), decoder_size))

    # Input feed: encoder inputs, decoder inputs, target_weights, as provided.
    input_feed = {}
    for l in xrange(encoder_size):
      input_feed[self.encoder_inputs[l].name] = encoder_inputs[l]
    for l in xrange(decoder_size):
      input_feed[self.decoder_inputs[l].name] = decoder_inputs[l]
      input_feed[self.target_weights[l].name] = target_weights[l]

    # Since our targets are decoder inputs shifted by one, we need one more.
    last_target = self.decoder_inputs[decoder_size].name
    input_feed[last_target] = np.zeros([self.batch_size], dtype=np.int32)

    # Output feed: depends on whether we do a backward step or not.
    if not forward_only:
      output_feed = [self.updates[bucket_id],  # Update Op that does SGD.
                     self.gradient_norms[bucket_id],  # Gradient norm.
                     self.losses[bucket_id]]  # Loss for this batch.
    else:
      output_feed = [self.losses[bucket_id]]  # Loss for this batch.
      for l in xrange(decoder_size):  # Output logits.
        output_feed.append(self.outputs[bucket_id][l])

    outputs = session.run(output_feed, input_feed)
    if not forward_only:
      return outputs[1], outputs[2], None  # Gradient norm, loss, no outputs.
    else:
      return None, outputs[0], outputs[1:]  # No gradient norm, loss, outputs.

  def get_batch(self, data, bucket_id):
    """Get a random batch of data from the specified bucket, prepare for step.

    To feed data in step(..) it must be a list of batch-major vectors, while
    data here contains single length-major cases. So the main logic of this
    function is to re-index data cases to be in the proper format for feeding.

    Args:
      data: a tuple of size len(self.buckets) in which each element contains
        lists of pairs of input and output data that we use to create a batch.
      bucket_id: integer, which bucket to get the batch for.

    Returns:
      The triple (encoder_inputs, decoder_inputs, target_weights) for
      the constructed batch that has the proper format to call step(...) later.
    """
    encoder_size, decoder_size = self.buckets[bucket_id]
    encoder_inputs, decoder_inputs = [], []

    # Get a random batch of encoder and decoder inputs from data,
    # pad them if needed, reverse encoder inputs and add GO to decoder.
    for _ in xrange(self.batch_size):
      encoder_input, decoder_input = random.choice(data[bucket_id])

      # Encoder inputs are padded and then reversed.
      encoder_pad = [data_utils.PAD_ID] * (encoder_size - len(encoder_input))
      encoder_inputs.append(list(reversed(encoder_input + encoder_pad)))

      # Decoder inputs get an extra "GO" symbol, and are padded then.
      decoder_pad_size = decoder_size - len(decoder_input) - 1
      decoder_inputs.append([data_utils.GO_ID] + decoder_input +
                            [data_utils.PAD_ID] * decoder_pad_size)

    # Now we create batch-major vectors from the data selected above.
    batch_encoder_inputs, batch_decoder_inputs, batch_weights = [], [], []

    # Batch encoder inputs are just re-indexed encoder_inputs.
    for length_idx in xrange(encoder_size):
      batch_encoder_inputs.append(
          np.array([encoder_inputs[batch_idx][length_idx]
                    for batch_idx in xrange(self.batch_size)], dtype=np.int32))

    # Batch decoder inputs are re-indexed decoder_inputs, we create weights.
    for length_idx in xrange(decoder_size):
      batch_decoder_inputs.append(
          np.array([decoder_inputs[batch_idx][length_idx]
                    for batch_idx in xrange(self.batch_size)], dtype=np.int32))

      # Create target_weights to be 0 for targets that are padding.
      batch_weight = np.ones(self.batch_size, dtype=np.float32)
      for batch_idx in xrange(self.batch_size):
        # We set weight to 0 if the corresponding target is a PAD symbol.
        # The corresponding target is decoder_input shifted by 1 forward.
        if length_idx < decoder_size - 1:
          target = decoder_inputs[batch_idx][length_idx + 1]
        if length_idx == decoder_size - 1 or target == data_utils.PAD_ID:
          batch_weight[batch_idx] = 0.0
      batch_weights.append(batch_weight)
    return batch_encoder_inputs, batch_decoder_inputs, batch_weights