Fix bug in evaluating FID and IS scores.

PiperOrigin-RevId: 340464023

Author: yan-wu

cs_gan/image_metrics.py

@@ -28,9 +28,9 @@ def get_image_metrics_for_samples(
   def sample_fn(arg):
     del arg
     samples = generator(prior.sample(max_classifier_batch))
-    # Ensure data is in [0, 1], as expected by TFGAN.
+    # Samples must be in [-1, 1], as expected by TFGAN.
     # Resizing to appropriate size is done by TFGAN.
-    return data_processor.postprocess(samples)
+    return samples
 
   fake_outputs = tfgan.eval.sample_and_run_inception(
       sample_fn,
@@ -39,7 +39,8 @@ def sample_fn(arg):
   fake_logits = fake_outputs['logits']
   inception_score = tfgan.eval.classifier_score_from_logits(fake_logits)
 
-  real_outputs = tfgan.eval.run_inception(real_images, num_batches=num_batches)
+  real_outputs = tfgan.eval.run_inception(
+      data_processor.preprocess(real_images), num_batches=num_batches)
   fid = tfgan.eval.frechet_classifier_distance_from_activations(
       real_outputs['pool_3'], fake_outputs['pool_3'])