Remove the 'cifar_unnormalized' dataset

build_label_map.py

@@ -113,7 +113,7 @@ def build_single_label_map(
             We may want many label_maps w/ same value of n_labeled.
             This allows us to disambiguate.
         dataset_name: A string representing the name of the dataset.
-            One of 'cifar10', 'svhn', 'imagenet', or 'cifar_unnormalized'.
+            One of 'cifar10', 'svhn', 'imagenet'.
         imagenet_path: A string that encodes the location of the raw imagenet
             data.
         fkeys_path: A string that encodes where to read fkeys from and write
@@ -143,7 +143,7 @@ def build_single_label_map(
             result_dict["values"] += random.sample(
                 unique_ids, n_labeled_per_class
             )
-    elif dataset_name in {"cifar10", "svhn", "cifar_unnormalized"}:
+    elif dataset_name in {"cifar10", "svhn"}:
         path = os.path.join(fkeys_path, dataset_name, "label_to_fkeys_train")
         with gfile.GFile(path, "r") as f:
             label_to_fkeys = json.load(f)

build_tfrecords.py

@@ -67,12 +67,6 @@
         "valid": 50050,
         "extra": 0,
     },
-    "cifar_unnormalized": {
-        "train": 50000,
-        "test": 10000,
-        "valid": 5000,
-        "extra": 0,
-    },
 }
 
 URLS = {
@@ -192,8 +186,6 @@ def main(_):
         train_set, test_set, extra_set = _load_svhn()
     elif FLAGS.dataset_name == "cifar10":
         train_set, test_set = _load_cifar10(normalize=True)
-    elif FLAGS.dataset_name == "cifar_unnormalized":
-        train_set, test_set = _load_cifar10(normalize=False)
     elif FLAGS.dataset_name == "imagenet_32":
         train_set, test_set = _load_imagenet_32()
     else:

lib/dataset_utils.py

@@ -39,42 +39,36 @@
 
 DATASET_SHAPE = {
     "cifar10": (None, 32, 32, 3),
-    "cifar_unnormalized": (None, 32, 32, 3),
     "svhn": (None, 32, 32, 3),
     "svhn_extra": (None, 32, 32, 3),
     "imagenet_32": (None, 32, 32, 3),
 }
 DATASET_DTYPE = {
     "cifar10": tf.float32,
-    "cifar_unnormalized": tf.uint8,
     "svhn": tf.uint8,
     "svhn_extra": tf.uint8,
     "imagenet_32": tf.uint8,
 }
 DATASET_CLASS_COUNT = {
     "cifar10": 10,
-    "cifar_unnormalized": 10,
     "svhn": 10,
     "svhn_extra": 10,
     "imagenet_32": 1000,
 }
 DATASET_EXAMPLE_COUNT = {
     "train": {
         "cifar10": 50000 - 5000,
-        "cifar_unnormalized": 50000 - 5000,
         "svhn": 73257 - 7326,
         "svhn_extra": 531131,
         "imagenet_32": 1281167 - 50050,
     },
     "test": {
         "cifar10": 10000,
-        "cifar_unnormalized": 10000,
         "svhn": 26032,
         "imagenet_32": 50000,
     },
     "valid": {
         "cifar10": 5000,
-        "cifar_unnormalized": 5000,
         "svhn": 7326,
         "imagenet_32": 50050,
     },
@@ -309,7 +303,6 @@ def get_filenames(dataset_name, split):
     if dataset_name in [
         "cifar10",
         "svhn",
-        "cifar_unnormalized",
         "imagenet_32",
     ]:
         filenames = [os.path.join(paths.TRAIN_DIR, dataset_name, FILES[split])]
@@ -332,8 +325,6 @@ def are_datasets_compatible(labeled_dataset_name, unlabeled_dataset_name):
     Boolean
   """
     valid_combos = [
-        ("cifar_unnormalized", "svhn"),
-        ("svhn", "cifar_unnormalized"),
         ("svhn", "svhn_extra"),
     ]
     return (labeled_dataset_name == unlabeled_dataset_name) or (

lib/hparams.py

@@ -75,7 +75,6 @@ def merge_dicts(x, y):
 
 # HParam overrides for different datasets
 cifar10_overrides = dict(horizontal_flip=True)
-cifar_unnormalized_overrides = cifar10_overrides
 
 imagenet_overrides = dict(
     horizontal_flip=True,
@@ -86,7 +85,6 @@ def merge_dicts(x, y):
 svhn_overrides = dict(gaussian_noise=False, vat_epsilon=1.0)
 dataset_overrides = dict(
     cifar10=cifar10_overrides,
-    cifar_unnormalized=cifar_unnormalized_overrides,
     imagenet=imagenet_overrides,
     imagenet_32=imagenet_overrides,
     imagenet_64=imagenet_overrides,