internal change

PiperOrigin-RevId: 219105444

compare_gan/src/multi_gan/README.md

@@ -1,21 +1,23 @@
-# A Case for Object Compositionality in Generative Models of Images
+# A Case for Object Compositionality in Deep Generative Models of Images
+
+![clevr-generated](illustrations/clevr_generated.png)
 
 This is the code repository complementing the
-["A Case for Object Compositionality in Generative Models of Images"](todo).
+["A Case for Object Compositionality in Deep Generative Models of Images"](https://arxiv.org/pdf/1810.10340.pdf).
 
 ## Datasets
 
-The following provides an overview of the datasets that were used.
+The following provides an overview of the datasets that were used. Corresponding
+.tfrecords files for all custom datasets are available [here](https://goo.gl/Eub81x).
 
 ### Multi-MNIST
 
-All Multi-MNIST datasets are available [here](todo). In these dataset each image
-consists of 3 (potentially overlapping) MNIST digits. Digits are obtained from
-the original dataset (using train/test/valid respectively), re-scaled and
-randomly placed in the image. A fixed-offset from the border ensures that digits
-appear entirely in the image.
+In these dataset each image consists of 3 (potentially overlapping) MNIST digits.
+Digits are obtained from the original dataset (using train/test/valid respectively),
+re-scaled and randomly placed in the image. A fixed-offset from the border ensures
+that digits appear entirely in the image.
 
-#### Uniform
+#### Independent
 
 All digits 0-9 have an equal chance of appearing in the image. This roughly
 results in a uniform distribution over all 3-tuples of digits in the images. In
@@ -41,7 +43,7 @@ Digits are sampled [uniformly](#uniform) and colored either
 appears in a Multi-MNIST image. Digits are drawn one by one into the canvas
 without blending colors, such that overlapping digits occlude one another.
 
-#### RGB Occluded + CIFAR10
+### CIFAR10 + RGB MM
 
 Draws digits from [rgb occluded](#rgb-occluded) images on top of a randomly
 sampled CIFAR10 image (resized to 64 x 64 using bilinear interpolation).

compare_gan/src/multi_gan/__init__.py

@@ -0,0 +1,19 @@
+# coding=utf-8
+# Copyright 2018 Google LLC & Hwalsuk Lee.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Init.py."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function

compare_gan/src/multi_gan/dataset.py

@@ -53,7 +53,7 @@ def unpack_clevr_image(image_data):
 def load_clevr(dataset_name, split_name, num_threads, buffer_size):
   del dataset_name
   filenames = tf.data.Dataset.list_files(
-      os.path.join(FLAGS.multigan_dataset_root, "clevr_%s*" % split_name))
+      os.path.join(FLAGS.multigan_dataset_root, "clevr/%s*" % split_name))
 
   return tf.data.TFRecordDataset(
       filenames,

compare_gan/src/multi_gan/visualize_gan.py

@@ -45,7 +45,8 @@
                     "Which checkpoint(s) to evaluate for a given study/task."
                     "Supports {'all', <int>}.")
 flags.DEFINE_enum("visualization_type", "multi_image", [
-    "multi_image", "latent", "multi_latent"], "How to visualize this GAN.")
+    "image", "multi_image", "latent", "multi_latent"],
+                  "How to visualize this GAN.")
 flags.DEFINE_integer("batch_size", 64, "Size of the batch.")
 flags.DEFINE_integer("images_per_fig", 4, "How many images to stack in a fig.")
 FLAGS = flags.FLAGS
@@ -206,6 +207,25 @@ def SaveMultiGanGeneratorImages(aggregated_images, generated_images, save_dir):
     plt.close(fig)
 
 
+def SaveGeneratorImages(aggregated_images, save_dir):
+  """Visualizes the aggregated output of all generators.
+
+  Args:
+    aggregated_images: The aggregated image (B, W, H, C).
+    save_dir: The path to the directory in which the figure should be saved.
+  """
+
+  n_images, _, _, _ = aggregated_images.shape
+
+  for i in range(n_images):
+    fig, ax = plt.subplots(figsize=(5, 5))
+
+    PlotImage(ax, aggregated_images[i])
+    plt.savefig(os.path.join(save_dir, "generator_images_%d.png" % i),
+                bbox_inches="tight")
+    plt.close(fig)
+
+
 def GetMultiGANGeneratorsOp(graph, gan_type, architecture, aggregate):
   """Returns the op to obtain the output of all generators."""
 
@@ -278,6 +298,16 @@ def EvalCheckpoint(checkpoint_path, task_workdir, options, out_cp_dir):
         fake_images, generator_preds = sess.run(fetches, feed_dict=feed_dict)
         SaveMultiGanGeneratorImages(fake_images, generator_preds, out_cp_dir)
 
+      # Compute outputs for GeneratorImages.
+      elif FLAGS.visualization_type == "image":
+        # Construct feed dict
+        z_sample = gan.z_generator(gan.batch_size, gan.z_dim)
+        feed_dict = {gan.z: z_sample}
+
+        # Fetch data and save images.
+        fake_images = sess.run(gan.fake_images, feed_dict=feed_dict)
+        SaveGeneratorImages(fake_images, out_cp_dir)
+
       # Compute outputs for MultiGanLatentTraversalImages
       elif (FLAGS.visualization_type == "multi_latent" and
             "MultiGAN" in options["gan_type"]):
@@ -362,9 +392,8 @@ def EvalTask(options, task_workdir, out_dir):
   if FLAGS.checkpoint == "all":
     all_checkpoint_paths = checkpoint_state.all_model_checkpoint_paths
   else:
-    all_checkpoint_paths = [
-        cp_path for cp_path in checkpoint_state.all_model_checkpoint_paths if
-        cp_path.split("-")[-1] == FLAGS.checkpoint]
+    all_checkpoint_paths = ["%s/checkpoint/%s.model-%s" % (
+        FLAGS.eval_task_workdir, options["gan_type"], FLAGS.checkpoint)]
 
   for checkpoint_path in all_checkpoint_paths:
     out_cp_dir = os.path.join(
@@ -405,6 +434,6 @@ def main(unused_argv):
   EvalTask(options, task_workdir, out_dir)
 
 if __name__ == "__main__":
-  flags.mark_flag_as_required("eval_workdir")
+  flags.mark_flag_as_required("eval_task_workdir")
   flags.mark_flag_as_required("out_dir")
   tf.app.run()