internal change

PiperOrigin-RevId: 204367849

README.md

@@ -5,6 +5,10 @@ paper (https://arxiv.org/abs/1711.10337) and in "The GAN Landscape: Losses, Arch
 
 If you want to see the version used only in the first paper - please see the *v1* branch of this repository.
 
+## Pre-trained models
+
+The pre-trained models are available on TensorFlow Hub. Please see [this colab](https://colab.research.google.com/github/google/compare_gan/blob/master/compare_gan/src/tfhub_models.ipynb)
+for an example how to use them.
 
 ### Best hyperparameters
 

compare_gan/bin/compare_gan_generate_tasks

@@ -14,6 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+#!/usr/bin/env python
+
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function

compare_gan/bin/compare_gan_run_one_task

@@ -14,6 +14,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+#!/usr/bin/env python
+
 from __future__ import absolute_import
 from __future__ import division
 

compare_gan/src/eval_gan_lib.py

@@ -398,16 +398,6 @@ def RunCheckpointEval(checkpoint_path, task_workdir, options, inception_graph):
   logging.info("Frechet Inception Distance for checkpoint %s is %.3f",
                checkpoint_path, result_dict["fid_score"])
 
-  # For comparison with other papers, for CIFAR dataset compute the FID score
-  # also on the whole training set.
-  if dataset == "cifar10":
-    logging.info("Computing FID score on 50k train set.")
-    train_images = GetRealImages(dataset, "train", 50000)
-    result_dict["fid50k_score"] = ComputeTFGanFIDScore(
-        fake_images, train_images, inception_graph)
-    logging.info("Frechet Inception Distance on 50k for checkpoint %s is %.3f",
-                 checkpoint_path, result_dict["fid50k_score"])
-
   if ShouldRunMultiscaleSSIM(options):
     result_dict["ms_ssim"] = ComputeMultiscaleSSIMScore(fake_images)
     logging.info("MS-SSIM score computed: %.3f", result_dict["ms_ssim"])
@@ -420,7 +410,7 @@ def RunTaskEval(options, task_workdir, inception_graph, out_file="scores.csv"):
   # If the output file doesn't exist, create it.
   csv_header = [
       "checkpoint_path", "model", "dataset", "tf_seed", "inception_score",
-      "fid_score", "fid50k_score", "ms_ssim_score", "train_accuracy",
+      "fid_score", "ms_ssim_score", "train_accuracy",
       "test_accuracy", "fake_accuracy", "train_d_loss", "test_d_loss",
       "sample_id"
   ]
@@ -464,7 +454,6 @@ def RunTaskEval(options, task_workdir, inception_graph, out_file="scores.csv"):
           checkpoint_path, options["gan_type"], options["dataset"], tf_seed,
           "%.3f" % result_dict.get("inception_score", default_value),
           "%.3f" % result_dict.get("fid_score", default_value),
-          "%.3f" % result_dict.get("fid50k_score", default_value),
           "%.3f" % result_dict.get("ms_ssim", default_value),
           "%.3f" % result_dict.get("train_accuracy", default_value),
           "%.3f" % result_dict.get("test_accuracy", default_value),

compare_gan/src/gan_lib_test.py

@@ -19,12 +19,23 @@
 from __future__ import division
 from __future__ import print_function
 
+import os
+import random
+import string
+
+from absl import flags
+from absl.testing import parameterized
+
 from compare_gan.src import gan_lib
+from compare_gan.src.gans import consts
 
+import numpy as np
 import tensorflow as tf
 
+FLAGS = flags.FLAGS
 
-class GanLibTest(tf.test.TestCase):
+
+class GanLibTest(parameterized.TestCase, tf.test.TestCase):
 
   def testLoadingTriangles(self):
     with tf.Graph().as_default():
@@ -60,6 +71,64 @@ def testLoadingMnist(self):
         self.assertEqual(image.shape, (28, 28, 1))
         self.assertEqual(label.shape, ())
 
+  def trainSingleStep(self, tf_seed):
+    """Train a GAN for a single training step and return the checkpoint."""
+    parameters = {
+        "tf_seed": tf_seed,
+        "learning_rate": 0.0002,
+        "z_dim": 64,
+        "batch_size": 2,
+        "training_steps": 1,
+        "disc_iters": 1,
+        "save_checkpoint_steps": 5000,
+        "discriminator_normalization": consts.NO_NORMALIZATION,
+        "dataset": "fake",
+        "gan_type": "GAN",
+        "penalty_type": consts.NO_PENALTY,
+        "architecture": consts.RESNET_CIFAR,
+        "lambda": 0.1,
+    }
+    random.seed(None)
+    exp_name = ''.join(random.choice(string.ascii_uppercase) for _ in range(16))
+    task_workdir = os.path.join(FLAGS.test_tmpdir, exp_name)
+    gan_lib.run_with_options(parameters, task_workdir)
+    ckpt_fn = os.path.join(task_workdir, "checkpoint/{}.model-0".format(
+                           parameters["gan_type"]))
+    tf.logging.info("ckpt_fn: %s", ckpt_fn)
+    self.assertTrue(tf.gfile.Exists(ckpt_fn + ".index"))
+    return tf.train.load_checkpoint(ckpt_fn)
+
+  def testSameTFRandomSeed(self):
+    # Setting the same tf_seed should give the same initial values at each run.
+    # In practice training still converge due to non-deterministic behavior
+    # in certain operations on the hardware level (e.g. cuDNN optimizations).
+    ckpt_1 = self.trainSingleStep(tf_seed=42)
+    ckpt_2 = self.trainSingleStep(tf_seed=42)
+
+    for name in ckpt_1.get_variable_to_shape_map():
+      self.assertTrue(ckpt_2.has_tensor(name))
+      t1 = ckpt_1.get_tensor(name)
+      t2 = ckpt_2.get_tensor(name)
+      np.testing.assert_almost_equal(t1, t2)
+
+  @parameterized.named_parameters([
+      ('Given', 1, 2),
+      ('OneNone', 1, None),
+      ('BothNone', None, None),
+  ])
+  def testDifferentTFRandomSeed(self, seed_1, seed_2):
+    ckpt_1 = self.trainSingleStep(tf_seed=seed_1)
+    ckpt_2 = self.trainSingleStep(tf_seed=seed_2)
+
+    diff_counter = 0
+    for name in ckpt_1.get_variable_to_shape_map():
+      self.assertTrue(ckpt_2.has_tensor(name))
+      t1 = ckpt_1.get_tensor(name)
+      t2 = ckpt_2.get_tensor(name)
+      if np.abs(t1 - t2).sum() > 0:
+         diff_counter += 1
+    self.assertGreater(diff_counter, 0)
+
 
 if __name__ == "__main__":
   tf.test.main()

compare_gan/src/gans/resnet_architecture.py

@@ -24,7 +24,6 @@
 
 from math import log
 
-# Dependency imports
 
 from compare_gan.src.gans import consts
 from compare_gan.src.gans import ops

compare_gan/src/generate_tasks_lib.py

@@ -442,9 +442,10 @@ def BestModelResnet19():
     # Line 3: FID score: 102.74
     {
       "dataset": "lsun-bedroom",
+      "gan_type": consts.LSGAN_WITH_PENALTY,
       "training_steps": 200000,
       "penalty_type": consts.NO_PENALTY,
-      "learning_rate": 0.000322,
+      "learning_rate": 0.0000322,
       "beta1": 0.5850,
       "beta2": 0.9904,
       "disc_iters": 1,
@@ -457,7 +458,7 @@ def BestModelResnet19():
       "dataset": "lsun-bedroom",
       "training_steps": 200000,
       "penalty_type": consts.NO_PENALTY,
-      "learning_rate": 0.000193,
+      "learning_rate": 0.0000193,
       "beta1": 0.1947,
       "beta2": 0.8819,
       "disc_iters": 1,
@@ -497,6 +498,7 @@ def BestModelResnet19():
     # Line 7: FID score: 41.6
     {
       "dataset": "lsun-bedroom",
+      "gan_type": consts.LSGAN_WITH_PENALTY,
       "training_steps": 200000,
       "penalty_type": consts.NO_PENALTY,
       "learning_rate": 0.0002,
@@ -580,11 +582,12 @@ def BestModelResnet19():
     model.update({
       "architecture": consts.RESNET5_ARCH,
       "batch_size": 64,
-      "gan_type": consts.GAN_WITH_PENALTY,
       "optimizer": "adam",
       "save_checkpoint_steps": 20000,
       "z_dim": 128,
     })
+    if "gan_type" not in model:
+      model.update({"gan_type": consts.GAN_WITH_PENALTY})
 
   return best_models
 
@@ -640,7 +643,7 @@ def BestModelResnetCifar():
       "beta1": 0.5,
       "beta2": 0.999,
       "disc_iters": 5,
-      "discriminator_normalization": consts.NO_NORMALIZATION,
+      "discriminator_normalization": consts.SPECTRAL_NORM,
       "tf_seed": 2,
       "lambda": 1,
     },