Fix compatibility for newer tensorflow and python3 (#6587)

1. In python3, dict.iteritems() is gone, change it to dict.items(). And it also doesn't treat zip object as a list, so change zip to list type.
2. Add encoding flag with open function to make it more compatible for Windows/Linux system.
3. For newer tensorflow, it expected int64 without giving a dtype parameter to embedding layer. We give it float32 to solve it.

research/adversarial_text/data/data_utils.py

@@ -325,8 +325,8 @@ def sort_vocab_by_frequency(vocab_freq_map):
 def write_vocab_and_frequency(ordered_vocab_freqs, output_dir):
   """Writes ordered_vocab_freqs into vocab.txt and vocab_freq.txt."""
   tf.gfile.MakeDirs(output_dir)
-  with open(os.path.join(output_dir, 'vocab.txt'), 'w') as vocab_f:
-    with open(os.path.join(output_dir, 'vocab_freq.txt'), 'w') as freq_f:
+  with open(os.path.join(output_dir, 'vocab.txt'), 'w', encoding='utf-8') as vocab_f:
+    with open(os.path.join(output_dir, 'vocab_freq.txt'), 'w', encoding='utf-8') as freq_f:
       for word, freq in ordered_vocab_freqs:
         vocab_f.write('{}\n'.format(word))
         freq_f.write('{}\n'.format(freq))

research/adversarial_text/data/document_generators.py

@@ -199,7 +199,7 @@ def imdb_documents(dataset='train',
       if is_validation and not include_validation:
         continue
 
-      with open(os.path.join(FLAGS.imdb_input_dir, d, filename)) as imdb_f:
+      with open(os.path.join(FLAGS.imdb_input_dir, d, filename), encoding='utf-8') as imdb_f:
         content = imdb_f.read()
       yield Document(
           content=content,
@@ -209,7 +209,7 @@ def imdb_documents(dataset='train',
           add_tokens=True)
 
   if FLAGS.amazon_unlabeled_input_file and include_unlabeled:
-    with open(FLAGS.amazon_unlabeled_input_file) as rt_f:
+    with open(FLAGS.amazon_unlabeled_input_file, encoding='utf-8') as rt_f:
       for content in rt_f:
         yield Document(
             content=content,

research/adversarial_text/gen_data.py

@@ -95,7 +95,7 @@ def make_vocab_ids(vocab_filename):
     ret[data.EOS_TOKEN] = len(string.printable)
     return ret
   else:
-    with open(vocab_filename) as vocab_f:
+    with open(vocab_filename, encoding='utf-8') as vocab_f:
       return dict([(line.strip(), i) for i, line in enumerate(vocab_f)])
 
 

research/adversarial_text/layers.py

@@ -67,7 +67,8 @@ class Embedding(K.layers.Layer):
       self.var = self.add_weight(
           shape=(self.vocab_size, self.embedding_dim),
           initializer=tf.random_uniform_initializer(-1., 1.),
-          name='embedding')
+          name='embedding',
+          dtype=tf.float32)
 
     if self.normalized:
       self.var = self._normalize(self.var)
@@ -152,7 +153,7 @@ class SoftmaxLoss(K.layers.Layer):
     self.multiclass_dense_layer = K.layers.Dense(self.vocab_size)
 
   def build(self, input_shape):
-    input_shape = input_shape[0]
+    input_shape = input_shape[0].as_list()
     with tf.device('/cpu:0'):
       self.lin_w = self.add_weight(
           shape=(input_shape[-1], self.vocab_size),
@@ -317,7 +318,7 @@ def optimize(loss,
     ne_grads, _ = tf.clip_by_global_norm(ne_grads, max_grad_norm)
     non_embedding_grads_and_vars = zip(ne_grads, ne_vars)
 
-    grads_and_vars = embedding_grads_and_vars + non_embedding_grads_and_vars
+    grads_and_vars = embedding_grads_and_vars + list(non_embedding_grads_and_vars)
 
     # Summarize
     _summarize_vars_and_grads(grads_and_vars)