TensorFlow 0.10 implementation of Variational Dropout (paper: https://arxiv.org/abs/1512.05287)

ptb_word_lm.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.
# ==============================================================================

"""Example / benchmark for building a PTB LSTM model.

Trains the model described in:
(Zaremba, et. al.) Recurrent Neural Network Regularization
http://arxiv.org/abs/1409.2329

There are 3 supported model configurations:
===========================================
| config | epochs | train | valid  | test
===========================================
| small  | 13     | 37.99 | 121.39 | 115.91
| medium | 39     | 48.45 |  86.16 |  82.07
| large  | 55     | 37.87 |  82.62 |  78.29
The exact results may vary depending on the random initialization.

The hyperparameters used in the model:
- init_scale - the initial scale of the weights
- learning_rate - the initial value of the learning rate
- max_grad_norm - the maximum permissible norm of the gradient
- num_layers - the number of LSTM layers
- num_steps - the number of unrolled steps of LSTM
- hidden_size - the number of LSTM units
- max_epoch - the number of epochs trained with the initial learning rate
- max_max_epoch - the total number of epochs for training
- keep_prob - the probability of keeping weights in the dropout layer
- lr_decay - the decay of the learning rate for each epoch after "max_epoch"
- batch_size - the batch size

The data required for this example is in the data/ dir of the
PTB dataset from Tomas Mikolov's webpage:

$ wget http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz
$ tar xvf simple-examples.tgz

To run:

$ python ptb_word_lm.py --data_path=simple-examples/data/

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

import time

import numpy as np
import tensorflow as tf

from tensorflow.models.rnn.ptb import reader

from variational_dropout_wrapper import VariationalDropoutWrapper, get_dropout_mask

flags = tf.flags
logging = tf.logging

flags.DEFINE_string(
    "model", "small",
    "A type of model. Possible options are: small, medium, large.")
flags.DEFINE_string("data_path", None, "data_path")
flags.DEFINE_bool("use_fp16", False,
                  "Train using 16-bit floats instead of 32bit floats")

FLAGS = flags.FLAGS


def data_type():
  return tf.float16 if FLAGS.use_fp16 else tf.float32


class PTBModel(object):
  """The PTB model."""

  def __init__(self, is_training, config):
    self.batch_size = batch_size = config.batch_size
    self.num_steps = num_steps = config.num_steps
    size = config.hidden_size
    vocab_size = config.vocab_size

    self._input_data = tf.placeholder(tf.int32, [batch_size, num_steps])
    self._targets = tf.placeholder(tf.int32, [batch_size, num_steps])

    # Slightly better results can be obtained with forget gate biases
    # initialized to 1 but the hyperparameters of the model would need to be
    # different than reported in the paper.
    lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(size, forget_bias=0.0)

    # To avoid using same dropout mask in different layers, create new dropout wrapper per layer
    cells = []
    for i in range(config.num_layers):
      with tf.variable_scope("layer" + i):
        cells.append(VariationalDropoutWrapper(lstm_cell, batch_size, keep_prob=config.keep_prob))
    cell = tf.nn.rnn_cell.MultiRNNCell(cells)

    self._initial_state = cell.zero_state(batch_size, data_type())

    with tf.device("/cpu:0"):
      embedding = tf.get_variable(
          "embedding", [vocab_size, size], dtype=data_type())
      inputs = tf.nn.embedding_lookup(embedding, self._input_data)

    if is_training and config.keep_prob < 1:
      # use same dropout mask across time steps, but keep different masks for samples and units
      dropout_mask = get_dropout_mask(config.keep_prob, [batch_size, size])
      inputs *= tf.expand_dims(dropout_mask, 1)

    # Simplified version of tensorflow.models.rnn.rnn.py's rnn().
    # This builds an unrolled LSTM for tutorial purposes only.
    # In general, use the rnn() or state_saving_rnn() from rnn.py.
    #
    # The alternative version of the code below is:
    #
    # inputs = [tf.squeeze(input_, [1])
    #           for input_ in tf.split(1, num_steps, inputs)]
    # outputs, state = tf.nn.rnn(cell, inputs, initial_state=self._initial_state)
    outputs = []
    state = self._initial_state
    with tf.variable_scope("RNN"):
      for time_step in range(num_steps):
        if time_step > 0: tf.get_variable_scope().reuse_variables()
        (cell_output, state) = cell(inputs[:, time_step, :], state)
        outputs.append(cell_output)

    output = tf.reshape(tf.concat(1, outputs), [-1, size])
    softmax_w = tf.get_variable(
        "softmax_w", [size, vocab_size], dtype=data_type())
    softmax_b = tf.get_variable("softmax_b", [vocab_size], dtype=data_type())
    logits = tf.matmul(output, softmax_w) + softmax_b
    loss = tf.nn.seq2seq.sequence_loss_by_example(
        [logits],
        [tf.reshape(self._targets, [-1])],
        [tf.ones([batch_size * num_steps], dtype=data_type())])
    self._cost = cost = tf.reduce_sum(loss) / batch_size
    self._final_state = state

    if not is_training:
      return

    self._lr = tf.Variable(0.0, trainable=False)
    tvars = tf.trainable_variables()
    grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars),
                                      config.max_grad_norm)
    optimizer = tf.train.GradientDescentOptimizer(self._lr)
    self._train_op = optimizer.apply_gradients(zip(grads, tvars))

    self._new_lr = tf.placeholder(
        tf.float32, shape=[], name="new_learning_rate")
    self._lr_update = tf.assign(self._lr, self._new_lr)

  def assign_lr(self, session, lr_value):
    session.run(self._lr_update, feed_dict={self._new_lr: lr_value})

  @property
  def input_data(self):
    return self._input_data

  @property
  def targets(self):
    return self._targets

  @property
  def initial_state(self):
    return self._initial_state

  @property
  def cost(self):
    return self._cost

  @property
  def final_state(self):
    return self._final_state

  @property
  def lr(self):
    return self._lr

  @property
  def train_op(self):
    return self._train_op


class SmallConfig(object):
  """Small config."""
  init_scale = 0.1
  learning_rate = 1.0
  max_grad_norm = 5
  num_layers = 2
  num_steps = 20
  hidden_size = 200
  max_epoch = 4
  max_max_epoch = 13
  keep_prob = 1.0
  lr_decay = 0.5
  batch_size = 20
  vocab_size = 10000


class MediumConfig(object):
  """Medium config."""
  init_scale = 0.05
  learning_rate = 1.0
  max_grad_norm = 5
  num_layers = 2
  num_steps = 35
  hidden_size = 650
  max_epoch = 6
  max_max_epoch = 39
  keep_prob = 0.5
  lr_decay = 0.8
  batch_size = 20
  vocab_size = 10000


class LargeConfig(object):
  """Large config."""
  init_scale = 0.04
  learning_rate = 1.0
  max_grad_norm = 10
  num_layers = 2
  num_steps = 35
  hidden_size = 1500
  max_epoch = 14
  max_max_epoch = 55
  keep_prob = 0.35
  lr_decay = 1 / 1.15
  batch_size = 20
  vocab_size = 10000


class TestConfig(object):
  """Tiny config, for testing."""
  init_scale = 0.1
  learning_rate = 1.0
  max_grad_norm = 1
  num_layers = 1
  num_steps = 2
  hidden_size = 2
  max_epoch = 1
  max_max_epoch = 1
  keep_prob = 1.0
  lr_decay = 0.5
  batch_size = 20
  vocab_size = 10000


def run_epoch(session, m, data, eval_op, verbose=False):
  """Runs the model on the given data."""
  epoch_size = ((len(data) // m.batch_size) - 1) // m.num_steps
  start_time = time.time()
  costs = 0.0
  iters = 0
  state = m.initial_state.eval()
  for step, (x, y) in enumerate(reader.ptb_iterator(data, m.batch_size,
                                                    m.num_steps)):
    cost, state, _ = session.run([m.cost, m.final_state, eval_op],
                                 {m.input_data: x,
                                  m.targets: y,
                                  m.initial_state: state})
    costs += cost
    iters += m.num_steps

    if verbose and step % (epoch_size // 10) == 10:
      print("%.3f perplexity: %.3f speed: %.0f wps" %
            (step * 1.0 / epoch_size, np.exp(costs / iters),
             iters * m.batch_size / (time.time() - start_time)))

  return np.exp(costs / iters)


def get_config():
  if FLAGS.model == "small":
    return SmallConfig()
  elif FLAGS.model == "medium":
    return MediumConfig()
  elif FLAGS.model == "large":
    return LargeConfig()
  elif FLAGS.model == "test":
    return TestConfig()
  else:
    raise ValueError("Invalid model: %s", FLAGS.model)


def main(_):
  if not FLAGS.data_path:
    raise ValueError("Must set --data_path to PTB data directory")

  raw_data = reader.ptb_raw_data(FLAGS.data_path)
  train_data, valid_data, test_data, _ = raw_data

  config = get_config()
  eval_config = get_config()
  eval_config.batch_size = 1
  eval_config.num_steps = 1

  with tf.Graph().as_default(), tf.Session() as session:
    initializer = tf.random_uniform_initializer(-config.init_scale,
                                                config.init_scale)
    with tf.variable_scope("model", reuse=None, initializer=initializer):
      m = PTBModel(is_training=True, config=config)
    with tf.variable_scope("model", reuse=True, initializer=initializer):
      mvalid = PTBModel(is_training=False, config=config)
      mtest = PTBModel(is_training=False, config=eval_config)

    tf.initialize_all_variables().run()

    for i in range(config.max_max_epoch):
      lr_decay = config.lr_decay ** max(i - config.max_epoch, 0.0)
      m.assign_lr(session, config.learning_rate * lr_decay)

      print("Epoch: %d Learning rate: %.3f" % (i + 1, session.run(m.lr)))
      train_perplexity = run_epoch(session, m, train_data, m.train_op,
                                   verbose=True)
      print("Epoch: %d Train Perplexity: %.3f" % (i + 1, train_perplexity))
      valid_perplexity = run_epoch(session, mvalid, valid_data, tf.no_op())
      print("Epoch: %d Valid Perplexity: %.3f" % (i + 1, valid_perplexity))

    test_perplexity = run_epoch(session, mtest, test_data, tf.no_op())
    print("Test Perplexity: %.3f" % test_perplexity)


if __name__ == "__main__":
  tf.app.run()

variational_dropout_wrapper.py

import tensorflow as tf
from tensorflow.python.ops.rnn_cell import RNNCell


def get_dropout_mask(keep_prob, shape):
  keep_prob = tf.convert_to_tensor(keep_prob)
  random_tensor = keep_prob + tf.random_uniform(shape)
  binary_tensor = tf.floor(random_tensor)
  dropout_mask = tf.inv(keep_prob) * binary_tensor
  return dropout_mask


class VariationalDropoutWrapper(RNNCell):
  def __init__(self, cell, batch_size, keep_prob):
    self._cell = cell
    self._output_dropout_mask = get_dropout_mask(keep_prob, [batch_size, cell.output_size])
    self._state_dropout_mask = get_dropout_mask(keep_prob, [batch_size, int(cell.state_size / 2)])

  @property
  def state_size(self):
    return self._cell.state_size

  @property
  def output_size(self):
    return self._cell.output_size

  def __call__(self, inputs, state, scope=None):
    # TODO: suppport non-LSTM cells and state_is_tuple=True
    c, h = tf.split(1, 2, state)
    h *= self._state_dropout_mask
    state = tf.concat(1, [c, h])
    output, new_state = self._cell(inputs, state, scope)
    output *= self._output_dropout_mask
    return output, new_state

LSTM language model performance on PennTreeBank dataset. Settings follow Zaremba's "medium" and Gal's untied/no MC version.

LSTM regularization	test perplexity
no dropout	114.5
all dropout (no shared mask)	108.4
input/output dropout（Zaremba 2014）	82.7
variational dropout（Gal 2016）	82.5

This is base on TensorFlow's RNN tutorial implementing Zaremba 2014.
https://github.com/tensorflow/tensorflow/blob/r0.10/tensorflow/models/rnn/ptb/ptb_word_lm.py

Is there any difference between the variational dropout here and dropout used by GAL of https://github.com/yaringal/DropoutUncertaintyExps/blob/master/YearPredictionMSD/net/net/net.py ?