Reproduce reported virtual adversarial text results

Author: Ryan Sepassi

adversarial_text/BUILD

@@ -1,10 +1,14 @@
+licenses(["notice"])  # Apache 2.0
+
 # Binaries
 # ==============================================================================
 py_binary(
     name = "evaluate",
     srcs = ["evaluate.py"],
     deps = [
         ":graphs",
+        # google3 file dep,
+        # tensorflow dep,
     ],
 )
 
@@ -14,6 +18,8 @@ py_binary(
     deps = [
         ":graphs",
         ":train_utils",
+        # google3 file dep,
+        # tensorflow dep,
     ],
 )
 
@@ -25,6 +31,8 @@ py_binary(
     deps = [
         ":graphs",
         ":train_utils",
+        # google3 file dep,
+        # tensorflow dep,
     ],
 )
 
@@ -37,30 +45,42 @@ py_library(
         ":adversarial_losses",
         ":inputs",
         ":layers",
+        # tensorflow dep,
     ],
 )
 
 py_library(
     name = "adversarial_losses",
     srcs = ["adversarial_losses.py"],
+    deps = [
+        # tensorflow dep,
+    ],
 )
 
 py_library(
     name = "inputs",
     srcs = ["inputs.py"],
     deps = [
+        # tensorflow dep,
         "//adversarial_text/data:data_utils",
     ],
 )
 
 py_library(
     name = "layers",
     srcs = ["layers.py"],
+    deps = [
+        # tensorflow dep,
+    ],
 )
 
 py_library(
     name = "train_utils",
     srcs = ["train_utils.py"],
+    deps = [
+        # numpy dep,
+        # tensorflow dep,
+    ],
 )
 
 # Tests
@@ -71,6 +91,7 @@ py_test(
     srcs = ["graphs_test.py"],
     deps = [
         ":graphs",
+        # tensorflow dep,
         "//adversarial_text/data:data_utils",
     ],
 )

adversarial_text/README.md

@@ -56,7 +56,6 @@ $ bazel run :pretrain -- \
     --embedding_dims=256 \
     --rnn_cell_size=1024 \
     --num_candidate_samples=1024 \
-    --optimizer=adam \
     --batch_size=256 \
     --learning_rate=0.001 \
     --learning_rate_decay_factor=0.9999 \
@@ -87,7 +86,6 @@ $ bazel run :train_classifier -- \
     --rnn_cell_size=1024 \
     --cl_num_layers=1 \
     --cl_hidden_size=30 \
-    --optimizer=adam \
     --batch_size=64 \
     --learning_rate=0.0005 \
     --learning_rate_decay_factor=0.9998 \
@@ -96,7 +94,8 @@ $ bazel run :train_classifier -- \
     --num_timesteps=400 \
     --keep_prob_emb=0.5 \
     --normalize_embeddings \
-    --adv_training_method=vat
+    --adv_training_method=vat \
+    --perturb_norm_length=5.0
 ```
 
 ### Evaluate on test data
@@ -136,21 +135,21 @@ adversarial training losses). The training loop itself is defined in
 ### Command-Line Flags
 
 Flags related to distributed training and the training loop itself are defined
-in `train_utils.py`.
+in [`train_utils.py`](https://github.com/tensorflow/models/tree/master/adversarial_text/train_utils.py).
 
-Flags related to model hyperparameters are defined in `graphs.py`.
+Flags related to model hyperparameters are defined in [`graphs.py`](https://github.com/tensorflow/models/tree/master/adversarial_text/graphs.py).
 
-Flags related to adversarial training are defined in `adversarial_losses.py`.
+Flags related to adversarial training are defined in [`adversarial_losses.py`](https://github.com/tensorflow/models/tree/master/adversarial_text/adversarial_losses.py).
 
 Flags particular to each job are defined in the main binary files.
 
 ### Data Generation
 
-*   Vocabulary generation: `gen_vocab.py`
-*   Data generation: `gen_data.py`
+*   Vocabulary generation: [`gen_vocab.py`](https://github.com/tensorflow/models/tree/master/adversarial_text/data/gen_vocab.py)
+*   Data generation: [`gen_data.py`](https://github.com/tensorflow/models/tree/master/adversarial_text/data/gen_data.py)
 
-Command-line flags defined in `document_generators.py` control which dataset is
-processed and how.
+Command-line flags defined in [`document_generators.py`](https://github.com/tensorflow/models/tree/master/adversarial_text/data/document_generators.py)
+control which dataset is processed and how.
 
 ## Contact for Issues
 

adversarial_text/adversarial_losses.py

@@ -1,4 +1,4 @@
-# Copyright 2017 Google, Inc. All Rights Reserved.
+# Copyright 2017 Google Inc. 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.
@@ -12,25 +12,27 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-
 """Adversarial losses for text models."""
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+# Dependency imports
+
 import tensorflow as tf
 
 flags = tf.app.flags
 FLAGS = flags.FLAGS
 
 # Adversarial and virtual adversarial training parameters.
-flags.DEFINE_float('perturb_norm_length', 0.1,
+flags.DEFINE_float('perturb_norm_length', 5.0,
                    'Norm length of adversarial perturbation to be '
-                   'optimized with validation')
+                   'optimized with validation. '
+                   '5.0 is optimal on IMDB with virtual adversarial training. ')
 
 # Virtual adversarial training parameters
 flags.DEFINE_integer('num_power_iteration', 1, 'The number of power iteration')
-flags.DEFINE_float('small_constant_for_finite_diff', 1e-3,
+flags.DEFINE_float('small_constant_for_finite_diff', 1e-1,
                    'Small constant for finite difference method')
 
 # Parameters for building the graph
@@ -83,19 +85,22 @@ def virtual_adversarial_loss(logits, embedded, inputs,
   """
   # Stop gradient of logits. See https://arxiv.org/abs/1507.00677 for details.
   logits = tf.stop_gradient(logits)
+
   # Only care about the KL divergence on the final timestep.
-  weights = _end_of_seq_mask(inputs.labels)
+  weights = inputs.eos_weights
+  assert weights is not None
 
   # Initialize perturbation with random noise.
   # shape(embedded) = (batch_size, num_timesteps, embedding_dim)
-  d = _mask_by_length(tf.random_normal(shape=tf.shape(embedded)), inputs.length)
+  d = tf.random_normal(shape=tf.shape(embedded))
 
   # Perform finite difference method and power iteration.
   # See Eq.(8) in the paper http://arxiv.org/pdf/1507.00677.pdf,
   # Adding small noise to input and taking gradient with respect to the noise
   # corresponds to 1 power iteration.
   for _ in xrange(FLAGS.num_power_iteration):
-    d = _scale_l2(d, FLAGS.small_constant_for_finite_diff)
+    d = _scale_l2(
+        _mask_by_length(d, inputs.length), FLAGS.small_constant_for_finite_diff)
     d_logits = logits_from_embedding_fn(embedded + d)
     kl = _kl_divergence_with_logits(logits, d_logits, weights)
     d, = tf.gradients(
@@ -104,8 +109,7 @@ def virtual_adversarial_loss(logits, embedded, inputs,
         aggregation_method=tf.AggregationMethod.EXPERIMENTAL_ACCUMULATE_N)
     d = tf.stop_gradient(d)
 
-  perturb = _scale_l2(
-      _mask_by_length(d, inputs.length), FLAGS.perturb_norm_length)
+  perturb = _scale_l2(d, FLAGS.perturb_norm_length)
   vadv_logits = logits_from_embedding_fn(embedded + perturb)
   return _kl_divergence_with_logits(logits, vadv_logits, weights)
 
@@ -136,7 +140,8 @@ def virtual_adversarial_loss_bidir(logits, embedded, inputs,
   """Virtual adversarial loss for bidirectional models."""
   logits = tf.stop_gradient(logits)
   f_inputs, _ = inputs
-  weights = _end_of_seq_mask(f_inputs.labels)
+  weights = f_inputs.eos_weights
+  assert weights is not None
 
   perturbs = [
       _mask_by_length(tf.random_normal(shape=tf.shape(emb)), f_inputs.length)
@@ -155,10 +160,7 @@ def virtual_adversarial_loss_bidir(logits, embedded, inputs,
         aggregation_method=tf.AggregationMethod.EXPERIMENTAL_ACCUMULATE_N)
     perturbs = [tf.stop_gradient(d) for d in perturbs]
 
-  perturbs = [
-      _scale_l2(_mask_by_length(d, f_inputs.length), FLAGS.perturb_norm_length)
-      for d in perturbs
-  ]
+  perturbs = [_scale_l2(d, FLAGS.perturb_norm_length) for d in perturbs]
   vadv_logits = logits_from_embedding_fn(
       [emb + d for (emb, d) in zip(embedded, perturbs)])
   return _kl_divergence_with_logits(logits, vadv_logits, weights)
@@ -167,40 +169,26 @@ def virtual_adversarial_loss_bidir(logits, embedded, inputs,
 def _mask_by_length(t, length):
   """Mask t, 3-D [batch, time, dim], by length, 1-D [batch,]."""
   maxlen = t.get_shape().as_list()[1]
-  mask = tf.sequence_mask(length, maxlen=maxlen)
+
+  # Subtract 1 from length to prevent the perturbation from going on 'eos'
+  mask = tf.sequence_mask(length - 1, maxlen=maxlen)
   mask = tf.expand_dims(tf.cast(mask, tf.float32), -1)
   # shape(mask) = (batch, num_timesteps, 1)
   return t * mask
 
 
 def _scale_l2(x, norm_length):
   # shape(x) = (batch, num_timesteps, d)
-
   # Divide x by max(abs(x)) for a numerically stable L2 norm.
   # 2norm(x) = a * 2norm(x/a)
   # Scale over the full sequence, dims (1, 2)
   alpha = tf.reduce_max(tf.abs(x), (1, 2), keep_dims=True) + 1e-12
-  l2_norm = alpha * tf.sqrt(tf.reduce_sum(tf.pow(x / alpha, 2), (1, 2),
-                                          keep_dims=True) + 1e-6)
+  l2_norm = alpha * tf.sqrt(
+      tf.reduce_sum(tf.pow(x / alpha, 2), (1, 2), keep_dims=True) + 1e-6)
   x_unit = x / l2_norm
   return norm_length * x_unit
 
 
-def _end_of_seq_mask(tokens):
-  """Generate a mask for the EOS token (1.0 on EOS, 0.0 otherwise).
-
-  Args:
-    tokens: 1-D integer tensor [num_timesteps*batch_size]. Each element is an
-      id from the vocab.
-
-  Returns:
-    Float tensor same shape as tokens, whose values are 1.0 on the end of
-    sequence and 0.0 on the others.
-  """
-  eos_id = FLAGS.vocab_size - 1
-  return tf.cast(tf.equal(tokens, eos_id), tf.float32)
-
-
 def _kl_divergence_with_logits(q_logits, p_logits, weights):
   """Returns weighted KL divergence between distributions q and p.
 
@@ -218,21 +206,19 @@ def _kl_divergence_with_logits(q_logits, p_logits, weights):
   # For logistic regression
   if FLAGS.num_classes == 2:
     q = tf.nn.sigmoid(q_logits)
-    p = tf.nn.sigmoid(p_logits)
     kl = (-tf.nn.sigmoid_cross_entropy_with_logits(logits=q_logits, labels=q) +
           tf.nn.sigmoid_cross_entropy_with_logits(logits=p_logits, labels=q))
+    kl = tf.squeeze(kl)
 
   # For softmax regression
   else:
-    q = tf.nn.softmax(q_logits)
-    p = tf.nn.softmax(p_logits)
-    kl = tf.reduce_sum(q * (tf.log(q) - tf.log(p)), 1)
+    kl = tf.reduce_sum(
+        q * (tf.nn.log_softmax(q_logits) - tf.nn.log_softmax(p_logits)), 1)
 
   num_labels = tf.reduce_sum(weights)
   num_labels = tf.where(tf.equal(num_labels, 0.), 1., num_labels)
 
-  kl.get_shape().assert_has_rank(2)
+  kl.get_shape().assert_has_rank(1)
   weights.get_shape().assert_has_rank(1)
-  loss = tf.identity(tf.reduce_sum(tf.expand_dims(weights, -1) * kl) /
-                     num_labels, name='kl')
+  loss = tf.identity(tf.reduce_sum(weights * kl) / num_labels, name='kl')
   return loss

adversarial_text/data/BUILD

@@ -1,3 +1,5 @@
+licenses(["notice"])  # Apache 2.0
+
 package(
     default_visibility = [
         "//adversarial_text:__subpackages__",
@@ -10,6 +12,7 @@ py_binary(
     deps = [
         ":data_utils",
         ":document_generators",
+        # tensorflow dep,
     ],
 )
 
@@ -19,23 +22,31 @@ py_binary(
     deps = [
         ":data_utils",
         ":document_generators",
+        # tensorflow dep,
     ],
 )
 
 py_library(
     name = "document_generators",
     srcs = ["document_generators.py"],
+    deps = [
+        # tensorflow dep,
+    ],
 )
 
 py_library(
     name = "data_utils",
     srcs = ["data_utils.py"],
+    deps = [
+        # tensorflow dep,
+    ],
 )
 
 py_test(
     name = "data_utils_test",
     srcs = ["data_utils_test.py"],
     deps = [
         ":data_utils",
+        # tensorflow dep,
     ],
 )

adversarial_text/data/data_utils.py

@@ -1,4 +1,4 @@
-# Copyright 2017 Google, Inc. All Rights Reserved.
+# Copyright 2017 Google Inc. 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.
@@ -12,13 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-
 """Utilities for generating/preprocessing data for adversarial text models."""
 
 import operator
 import os
 import random
 import re
+
+# Dependency imports
+
 import tensorflow as tf
 
 EOS_TOKEN = '</s>'
@@ -215,13 +217,17 @@ def build_lm_sequence(seq):
 
   Returns:
     SequenceWrapper with `seq` tokens copied over to output sequence tokens and
-    labels (offset by 1, i.e. predict next token) with weights set to 1.0.
+    labels (offset by 1, i.e. predict next token) with weights set to 1.0,
+    except for <eos> token.
   """
   lm_seq = SequenceWrapper()
-  for i, timestep in enumerate(seq[:-1]):
-    lm_seq.add_timestep().set_token(timestep.token).set_label(
-        seq[i + 1].token).set_weight(1.0)
-
+  for i, timestep in enumerate(seq):
+    if i == len(seq) - 1:
+      lm_seq.add_timestep().set_token(timestep.token).set_label(
+          seq[i].token).set_weight(0.0)
+    else:
+      lm_seq.add_timestep().set_token(timestep.token).set_label(
+          seq[i + 1].token).set_weight(1.0)
   return lm_seq