Merge pull request #646 from rsepassi/push

v1.5.5

Author: lukaszkaiser

.travis.yml

@@ -16,7 +16,7 @@ matrix:
     - python: "3.6"
       env: TF_VERSION="1.4.*"
     - python: "3.6"
-      env: TF_VERSION="1.6.*"
+      env: TF_VERSION="1.5.*"
 before_install:
   - echo "deb [arch=amd64] http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal" | sudo tee /etc/apt/sources.list.d/tensorflow-serving.list
   - curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | sudo apt-key add -

.github/ISSUE_TEMPLATE.md renamed to ISSUE_TEMPLATE.md

@@ -6,7 +6,7 @@
 
 ### *TensorFlow* and *tensor2tensor* versions
 
-<!-- **Note** Run `pip list | grep tensor` to include TensorFlow and tensor2tensor versions -->
+<!-- **Note** Run `pip freeze | grep tensor` to get versions -->
 
 > …
 
@@ -16,7 +16,7 @@
 
 ### In case of bug report: Error log
 
-<!-- Please use code markdown to format output messages. -->
+<!-- Please use code markdown (```) to format output messages. -->
 <!-- See https://help.github.com/articles/creating-and-highlighting-code-blocks/ -->
 
 > …

README.md

@@ -15,7 +15,7 @@ welcome](https://img.shields.io/badge/contributions-welcome-brightgreen.svg)](CO
 of deep learning models and datasets designed to make deep learning more
 accessible and [accelerate ML
 research](https://research.googleblog.com/2017/06/accelerating-deep-learning-research.html).
- is actively used and maintained by researchers and engineers within the
+T2T is actively used and maintained by researchers and engineers within the
 [Google Brain team](https://research.google.com/teams/brain/) and a community
 of users. We're eager to collaborate with you too, so feel free to
 [open an issue on GitHub](https://github.com/tensorflow/tensor2tensor/issues)

docs/cloud_tpu.md

@@ -18,6 +18,17 @@ See the official tutorial for [running Transfomer
 on Cloud TPUs](https://cloud.google.com/tpu/docs/tutorials/transformer)
 for some examples and try out your own problems.
 
+Image Transformer:
+* `imagetransformer` with `imagetransformer_base_tpu` (or
+    `imagetransformer_tiny_tpu`)
+* `img2img_transformer` with `img2img_transformer_base_tpu` (or
+    `img2img_transformer_tiny_tpu`)
+
+You can run the `ImageTransformer` model on problems like unconditional or
+conditional Image generation and `Img2ImgTransformer` model on Super Resolution.
+We run on datasets like CelebA, CIFAR and ImageNet but they should work with any
+other image dataset.
+
 Residual networks:
 * `resnet` with `resnet_50` (or `resnet_18` or `resnet_34`)
 * `revnet` with `revnet_104` (or `revnet_38_cifar`)

docs/walkthrough.md

@@ -15,7 +15,7 @@ welcome](https://img.shields.io/badge/contributions-welcome-brightgreen.svg)](CO
 of deep learning models and datasets designed to make deep learning more
 accessible and [accelerate ML
 research](https://research.googleblog.com/2017/06/accelerating-deep-learning-research.html).
- is actively used and maintained by researchers and engineers within the
+T2T is actively used and maintained by researchers and engineers within the
 [Google Brain team](https://research.google.com/teams/brain/) and a community
 of users. We're eager to collaborate with you too, so feel free to
 [open an issue on GitHub](https://github.com/tensorflow/tensor2tensor/issues)
@@ -154,7 +154,7 @@ For all translation problems, we suggest to try the Transformer model:
 this should reach a BLEU score of about 28 on the English-German data-set,
 which is close to state-of-the art. If training on a single GPU, try the
 `--hparams_set=transformer_base_single_gpu` setting. For very good results
-or larger data-sets (e.g., for English-French)m, try the big model
+or larger data-sets (e.g., for English-French), try the big model
 with `--hparams_set=transformer_big`.
 
 ## Basics

setup.py

@@ -5,7 +5,7 @@
 
 setup(
     name='tensor2tensor',
-    version='1.5.4',
+    version='1.5.5',
     description='Tensor2Tensor',
     author='Google Inc.',
     author_email='[email protected]',

tensor2tensor/data_generators/cifar.py

@@ -124,7 +124,8 @@ def preprocess_example(self, example, mode, unused_hparams):
     image.set_shape([_CIFAR10_IMAGE_SIZE, _CIFAR10_IMAGE_SIZE, 3])
     if mode == tf.estimator.ModeKeys.TRAIN:
       image = image_utils.cifar_image_augmentation(image)
-    image = tf.image.per_image_standardization(image)
+    if not self._was_reversed:
+      image = tf.image.per_image_standardization(image)
     example["inputs"] = image
     return example
 
@@ -151,7 +152,8 @@ class ImageCifar10Plain(ImageCifar10):
   def preprocess_example(self, example, mode, unused_hparams):
     image = example["inputs"]
     image.set_shape([_CIFAR10_IMAGE_SIZE, _CIFAR10_IMAGE_SIZE, 3])
-    image = tf.image.per_image_standardization(image)
+    if not self._was_reversed:
+      image = tf.image.per_image_standardization(image)
     example["inputs"] = image
     return example
 
@@ -179,7 +181,8 @@ def dataset_filename(self):
   def preprocess_example(self, example, mode, unused_hparams):
     image = example["inputs"]
     image = image_utils.resize_by_area(image, 8)
-    image = tf.image.per_image_standardization(image)
+    if not self._was_reversed:
+      image = tf.image.per_image_standardization(image)
     example["inputs"] = image
     return example
 
@@ -192,7 +195,6 @@ def dataset_filename(self):
     return "image_cifar10_plain"  # Reuse CIFAR-10 plain data.
 
   def preprocess_example(self, example, unused_mode, unused_hparams):
-
     inputs = example["inputs"]
     # For Img2Img resize input and output images as desired.
     example["inputs"] = image_utils.resize_by_area(inputs, 8)
@@ -330,7 +332,8 @@ def preprocess_example(self, example, mode, unused_hparams):
     image.set_shape([_CIFAR100_IMAGE_SIZE, _CIFAR100_IMAGE_SIZE, 3])
     if mode == tf.estimator.ModeKeys.TRAIN:
       image = image_utils.cifar_image_augmentation(image)
-    image = tf.image.per_image_standardization(image)
+    if not self._was_reversed:
+      image = tf.image.per_image_standardization(image)
     example["inputs"] = image
     return example
 
@@ -357,7 +360,8 @@ class ImageCifar100Plain(ImageCifar100):
   def preprocess_example(self, example, mode, unused_hparams):
     image = example["inputs"]
     image.set_shape([_CIFAR100_IMAGE_SIZE, _CIFAR100_IMAGE_SIZE, 3])
-    image = tf.image.per_image_standardization(image)
+    if not self._was_reversed:
+      image = tf.image.per_image_standardization(image)
     example["inputs"] = image
     return example
 
@@ -385,7 +389,8 @@ def dataset_filename(self):
   def preprocess_example(self, example, mode, unused_hparams):
     image = example["inputs"]
     image = image_utils.resize_by_area(image, 8)
-    image = tf.image.per_image_standardization(image)
+    if not self._was_reversed:
+      image = tf.image.per_image_standardization(image)
     example["inputs"] = image
     return example
 
@@ -398,7 +403,6 @@ def dataset_filename(self):
     return "image_cifar100_plain"  # Reuse CIFAR-100 plain data.
 
   def preprocess_example(self, example, unused_mode, unused_hparams):
-
     inputs = example["inputs"]
     # For Img2Img resize input and output images as desired.
     example["inputs"] = image_utils.resize_by_area(inputs, 8)

tensor2tensor/data_generators/gym.py

@@ -35,6 +35,8 @@
 import tensorflow as tf
 
 
+
+
 flags = tf.flags
 FLAGS = flags.FLAGS
 
@@ -157,7 +159,6 @@ def num_steps(self):
     return 5000
 
 
-
 @registry.register_problem
 class GymPongTrajectoriesFromPolicy(GymDiscreteProblem):
   """Pong game, loaded actions."""
@@ -197,7 +198,7 @@ def generator(self, data_dir, tmp_dir):
         model_saver.restore(sess, FLAGS.model_path)
         for item in super(GymPongTrajectoriesFromPolicy,
                           self).generator(data_dir, tmp_dir):
-            yield item
+          yield item
 
   # TODO(blazej0): For training of atari agents wrappers are usually used.
   # Below we have a hacky solution which is a workaround to be used together

tensor2tensor/data_generators/image_utils.py

@@ -26,6 +26,7 @@
 from tensor2tensor.data_generators import generator_utils
 from tensor2tensor.data_generators import problem
 from tensor2tensor.data_generators import text_encoder
+from tensor2tensor.utils import metrics
 from tensor2tensor.utils import registry
 
 import tensorflow as tf
@@ -64,9 +65,19 @@ def example_reading_spec(self, label_repr=None):
     return data_fields, data_items_to_decoders
 
   def preprocess_example(self, example, mode, hparams):
-    example["inputs"] = tf.image.per_image_standardization(example["inputs"])
+    if not self._was_reversed:
+      example["inputs"] = tf.image.per_image_standardization(example["inputs"])
     return example
 
+  def eval_metrics(self):
+    eval_metrics = [
+        metrics.Metrics.ACC, metrics.Metrics.ACC_TOP5,
+        metrics.Metrics.ACC_PER_SEQ, metrics.Metrics.NEG_LOG_PERPLEXITY
+    ]
+    if self._was_reversed:
+      eval_metrics += [metrics.Metrics.IMAGE_SUMMARY]
+    return eval_metrics
+
 
 class Image2ClassProblem(ImageProblem):
   """Base class for image classification problems."""

tensor2tensor/data_generators/imagenet.py

@@ -334,7 +334,8 @@ def distorted_bounding_box_crop(image,
   Returns:
     (cropped image `Tensor`, distorted bbox `Tensor`).
   """
-  with tf.name_scope(scope, default_name="distorted_bounding_box_crop", values=[image, bbox]):
+  with tf.name_scope(scope, default_name="distorted_bounding_box_crop",
+                     values=[image, bbox]):
     # Each bounding box has shape [1, num_boxes, box coords] and
     # the coordinates are ordered [ymin, xmin, ymax, xmax].