Fix github issue #3269 where the accuracy is wrongly underestimated for binary classification and build issue #2784

research/adversarial_text/README.md

@@ -4,7 +4,6 @@ Code for [*Adversarial Training Methods for Semi-Supervised Text Classification*
 
 ## Requirements
 
-* Bazel ([install](https://bazel.build/versions/master/docs/install.html))
 * TensorFlow >= v1.1
 
 ## End-to-end IMDB Sentiment Classification
@@ -23,7 +22,7 @@ The directory `/tmp/aclImdb` contains the raw IMDB data.
 
 ```
 $ IMDB_DATA_DIR=/tmp/imdb
-$ bazel run data:gen_vocab -- \
+$ python gen_vocab.py -- \
     --output_dir=$IMDB_DATA_DIR \
     --dataset=imdb \
     --imdb_input_dir=/tmp/aclImdb \
@@ -35,7 +34,7 @@ Vocabulary and frequency files will be generated in `$IMDB_DATA_DIR`.
 ###  Generate training, validation, and test data
 
 ```
-$ bazel run data:gen_data -- \
+$ python gen_data.py -- \
     --output_dir=$IMDB_DATA_DIR \
     --dataset=imdb \
     --imdb_input_dir=/tmp/aclImdb \
@@ -49,7 +48,7 @@ $ bazel run data:gen_data -- \
 
 ```
 $ PRETRAIN_DIR=/tmp/models/imdb_pretrain
-$ bazel run :pretrain -- \
+$ python pretrain.py -- \
     --train_dir=$PRETRAIN_DIR \
     --data_dir=$IMDB_DATA_DIR \
     --vocab_size=86934 \
@@ -77,7 +76,7 @@ training and classification.
 
 ```
 $ TRAIN_DIR=/tmp/models/imdb_classify
-$ bazel run :train_classifier -- \
+$ python train_classifier.py -- \
     --train_dir=$TRAIN_DIR \
     --pretrained_model_dir=$PRETRAIN_DIR \
     --data_dir=$IMDB_DATA_DIR \
@@ -102,7 +101,7 @@ $ bazel run :train_classifier -- \
 
 ```
 $ EVAL_DIR=/tmp/models/imdb_eval
-$ bazel run :evaluate -- \
+$ python evaluate.py -- \
     --eval_dir=$EVAL_DIR \
     --checkpoint_dir=$TRAIN_DIR \
     --eval_data=test \
@@ -145,13 +144,13 @@ Flags particular to each job are defined in the main binary files.
 
 ### Data Generation
 
-*   Vocabulary generation: [`gen_vocab.py`](https://github.com/tensorflow/models/tree/master/research/adversarial_text/data/gen_vocab.py)
-*   Data generation: [`gen_data.py`](https://github.com/tensorflow/models/tree/master/research/adversarial_text/data/gen_data.py)
+*   Vocabulary generation: [`gen_vocab.py`](https://github.com/tensorflow/models/tree/master/research/adversarial_text/gen_vocab.py)
+*   Data generation: [`gen_data.py`](https://github.com/tensorflow/models/tree/master/research/adversarial_text/gen_data.py)
 
 Command-line flags defined in [`document_generators.py`](https://github.com/tensorflow/models/tree/master/research/adversarial_text/data/document_generators.py)
 control which dataset is processed and how.
 
 ## Contact for Issues
 
 * Ryan Sepassi, @rsepassi
-* Andrew M. Dai, @a-dai
+* Andrew M. Dai, @a-dai <adai@google.com>

research/adversarial_text/data/data_utils_test.py

@@ -21,7 +21,7 @@
 
 import tensorflow as tf
 
-from adversarial_text.data import data_utils
+from data import data_utils
 
 data = data_utils
 

research/adversarial_text/data/document_generators.py

@@ -26,7 +26,7 @@
 
 import tensorflow as tf
 
-from adversarial_text.data import data_utils
+from data import data_utils
 
 flags = tf.app.flags
 FLAGS = flags.FLAGS

research/adversarial_text/evaluate.py

@@ -25,7 +25,7 @@
 
 import tensorflow as tf
 
-from adversarial_text import graphs
+import graphs
 
 flags = tf.app.flags
 FLAGS = flags.FLAGS

research/adversarial_text/data/gen_data.py → research/adversarial_text/gen_data.py

@@ -34,8 +34,8 @@
 
 import tensorflow as tf
 
-from adversarial_text.data import data_utils
-from adversarial_text.data import document_generators
+from data import data_utils
+from data import document_generators
 
 data = data_utils
 flags = tf.app.flags

research/adversarial_text/data/gen_vocab.py → research/adversarial_text/gen_vocab.py

@@ -23,8 +23,8 @@
 
 import tensorflow as tf
 
-from adversarial_text.data import data_utils
-from adversarial_text.data import document_generators
+from data import data_utils
+from data import document_generators
 
 flags = tf.app.flags
 FLAGS = flags.FLAGS

research/adversarial_text/graphs.py

@@ -24,9 +24,9 @@
 
 import tensorflow as tf
 
-from adversarial_text import adversarial_losses as adv_lib
-from adversarial_text import inputs as inputs_lib
-from adversarial_text import layers as layers_lib
+import adversarial_losses as adv_lib
+import inputs as inputs_lib
+import layers as layers_lib
 
 flags = tf.app.flags
 FLAGS = flags.FLAGS

research/adversarial_text/graphs_test.py

@@ -29,8 +29,8 @@
 
 import tensorflow as tf
 
-from adversarial_text import graphs
-from adversarial_text.data import data_utils
+import graphs
+from data import data_utils
 
 flags = tf.app.flags
 FLAGS = flags.FLAGS

research/adversarial_text/inputs.py

@@ -24,7 +24,7 @@
 
 import tensorflow as tf
 
-from adversarial_text.data import data_utils
+from data import data_utils
 
 
 class VatxtInput(object):

research/adversarial_text/layers.py

@@ -254,7 +254,7 @@ def predictions(logits):
   with tf.name_scope('predictions'):
     # For binary classification
     if inner_dim == 1:
-      pred = tf.cast(tf.greater(tf.squeeze(logits, -1), 0.5), tf.int64)
+      pred = tf.cast(tf.greater(tf.squeeze(logits, -1), 0.), tf.int64)
     # For multi-class classification
     else:
       pred = tf.argmax(logits, 2)

research/adversarial_text/pretrain.py

@@ -27,8 +27,8 @@
 
 import tensorflow as tf
 
-from adversarial_text import graphs
-from adversarial_text import train_utils
+import graphs
+import train_utils
 
 FLAGS = tf.app.flags.FLAGS
 

research/adversarial_text/train_classifier.py

@@ -35,8 +35,8 @@
 
 import tensorflow as tf
 
-from adversarial_text import graphs
-from adversarial_text import train_utils
+import graphs
+import train_utils
 
 flags = tf.app.flags
 FLAGS = flags.FLAGS