T2T depends on TF 1.4+, daisy_chain_getter bug fix, some Eager-mode improvements/fixes

PiperOrigin-RevId: 177538074

Author: Ryan Sepassi

docs/example_life.md

@@ -75,7 +75,7 @@ hooks in the `Problem` class and the model's `HParams` object (typically
 registered in the model's file and specified by the `--hparams_set` flag).
 
 The entire input pipeline is implemented with the new `tf.data.Dataset` API
-(previously `tf.contrib.data.Dataset`).
+(previously `tf.data.Dataset`).
 
 The key function in the codebase for the input pipeline is
 [`data_reader.input_pipeline`](https://github.com/tensorflow/tensor2tensor/tree/master/tensor2tensor/utils/data_reader.py).

setup.py

@@ -15,8 +15,7 @@
     package_data={
         'tensor2tensor.data_generators': ['test_data/*'],
         'tensor2tensor.visualization': [
-            'attention.js',
-            'TransformerVisualization.ipynb'
+            'attention.js', 'TransformerVisualization.ipynb'
         ],
     },
     scripts=[
@@ -34,8 +33,8 @@
         'six',
     ],
     extras_require={
-        'tensorflow': ['tensorflow>=1.3.0'],
-        'tensorflow_gpu': ['tensorflow-gpu>=1.3.0'],
+        'tensorflow': ['tensorflow>=1.4.0'],
+        'tensorflow_gpu': ['tensorflow-gpu>=1.4.0'],
         'tests': ['pytest', 'h5py', 'mock'],
     },
     classifiers=[
@@ -45,4 +44,5 @@
         'License :: OSI Approved :: Apache Software License',
         'Topic :: Scientific/Engineering :: Artificial Intelligence',
     ],
-    keywords='tensorflow machine learning',)
+    keywords='tensorflow machine learning',
+)

tensor2tensor/data_generators/generator_utils.py

@@ -125,6 +125,13 @@ def shard_filepath(fname, num_shards):
   ]
 
 
+def outputs_exist(filenames):
+  for out_fname in filenames:
+    out_fname = out_fname.replace(UNSHUFFLED_SUFFIX, "")
+    if tf.gfile.Exists(out_fname):
+      return out_fname
+
+
 def generate_files(generator, output_filenames, max_cases=None):
   """Generate cases from a generator and save as TFRecord files.
 
@@ -137,6 +144,9 @@ def generate_files(generator, output_filenames, max_cases=None):
     max_cases: maximum number of cases to get from the generator;
       if None (default), we use the generator until StopIteration is raised.
   """
+  if outputs_exist(output_filenames):
+    tf.logging.info("Skipping generator because outputs files exist")
+    return
   num_shards = len(output_filenames)
   writers = [tf.python_io.TFRecordWriter(fname) for fname in output_filenames]
   counter, shard = 0, 0
@@ -440,6 +450,9 @@ def generate_dataset_and_shuffle(train_gen,
 
 
 def shuffle_dataset(filenames):
+  if outputs_exist(filenames):
+    tf.logging.info("Skipping shuffle because output files exist")
+    return
   tf.logging.info("Shuffling data...")
   for fname in filenames:
     records = read_records(fname)

tensor2tensor/data_generators/image.py

@@ -42,6 +42,8 @@
 
 import tensorflow as tf
 
+from tensorflow.python.eager import context
+
 
 def resize_by_area(img, size):
   """image resize function used by quite a few image problems."""
@@ -463,6 +465,21 @@ def hparams(self, defaults, unused_model_hparams):
     p.target_space_id = 1
 
 
+def _encoded_images(images):
+  if context.in_eager_mode():
+    for image in images:
+      yield tf.image.encode_png(image).numpy()
+  else:
+    (width, height, channels) = images[0].shape
+    with tf.Graph().as_default():
+      image_t = tf.placeholder(dtype=tf.uint8, shape=(width, height, channels))
+      encoded_image_t = tf.image.encode_png(image_t)
+      with tf.Session() as sess:
+        for image in images:
+          enc_string = sess.run(encoded_image_t, feed_dict={image_t: image})
+          yield enc_string
+
+
 def image_generator(images, labels):
   """Generator for images that takes image and labels lists and creates pngs.
 
@@ -484,20 +501,15 @@ def image_generator(images, labels):
   """
   if not images:
     raise ValueError("Must provide some images for the generator.")
-  (width, height, channels) = images[0].shape
-  with tf.Graph().as_default():
-    image_t = tf.placeholder(dtype=tf.uint8, shape=(width, height, channels))
-    encoded_image_t = tf.image.encode_png(image_t)
-    with tf.Session() as sess:
-      for (image, label) in zip(images, labels):
-        enc_string = sess.run(encoded_image_t, feed_dict={image_t: image})
-        yield {
-            "image/encoded": [enc_string],
-            "image/format": ["png"],
-            "image/class/label": [int(label)],
-            "image/height": [height],
-            "image/width": [width]
-        }
+  width, height, _ = images[0].shape
+  for (enc_image, label) in zip(_encoded_images(images), labels):
+    yield {
+        "image/encoded": [enc_image],
+        "image/format": ["png"],
+        "image/class/label": [int(label)],
+        "image/height": [height],
+        "image/width": [width]
+    }
 
 
 # URLs and filenames for MNIST data.

tensor2tensor/data_generators/problem.py

@@ -382,7 +382,7 @@ def dataset(self,
         data_filepattern)
     if shuffle_files or shuffle_files is None and is_training:
       random.shuffle(data_files)
-    dataset = tf.contrib.data.TFRecordDataset(data_files)
+    dataset = tf.data.TFRecordDataset(data_files)
 
     def decode_record(record):
       """Serialized Example to dict of <feature name, Tensor>."""
@@ -399,13 +399,12 @@ def _preprocess(example):
       self.maybe_copy_features(example)
       return example
 
-    dataset = dataset.map(decode_record, num_threads=num_threads)
+    dataset = dataset.map(decode_record, num_parallel_calls=num_threads)
 
     if preprocess:
-      dataset = dataset.map(
-          _preprocess,
-          num_threads=num_threads,
-          output_buffer_size=output_buffer_size)
+      dataset = dataset.map(_preprocess, num_parallel_calls=num_threads)
+    if output_buffer_size:
+      dataset = dataset.prefetch(output_buffer_size)
 
     return dataset
 
@@ -517,7 +516,7 @@ def define_shapes(example):
     dataset = self.dataset(
         mode=mode, data_dir=data_dir, num_threads=num_threads, hparams=hparams)
     dataset = dataset.map(
-        data_reader.cast_int64_to_int32, num_threads=num_threads)
+        data_reader.cast_int64_to_int32, num_parallel_calls=num_threads)
     if is_training:
       dataset = dataset.repeat(None)
 

tensor2tensor/layers/rev_block.py

@@ -399,7 +399,7 @@ def grad_fn(inputs, variables, outputs, output_grads):
   @common_layers.fn_with_custom_grad(grad_fn)
   def fn_with_recompute(*args):
     cached_vs.append(tf.get_variable_scope())
-    # TODO(rsepassi): Rm conditional in TF 1.4
+    # TODO(rsepassi): Rm conditional in TF 1.5
     if hasattr(tf.contrib.framework, "current_arg_scope"):
       cached_arg_scope.append(tf.contrib.framework.current_arg_scope())
     else: