@@ -77,12 +77,17 @@ def cifar_generator(cifar_version, tmp_dir, training, how_many, start_from=0):
7777 test_files = _CIFAR10_TEST_FILES
7878 prefix = _CIFAR10_PREFIX
7979 image_size = _CIFAR10_IMAGE_SIZE
80- elif cifar_version == "cifar100" :
80+ label_key = "labels"
81+ elif cifar_version == "cifar100" or cifar_version == "cifar20" :
8182 url = _CIFAR100_URL
8283 train_files = _CIFAR100_TRAIN_FILES
8384 test_files = _CIFAR100_TEST_FILES
8485 prefix = _CIFAR100_PREFIX
8586 image_size = _CIFAR100_IMAGE_SIZE
87+ if cifar_version == "cifar100" :
88+ label_key = "fine_labels"
89+ else :
90+ label_key = "coarse_labels"
8691
8792 _get_cifar (tmp_dir , url )
8893 data_files = train_files if training else test_files
@@ -97,7 +102,7 @@ def cifar_generator(cifar_version, tmp_dir, training, how_many, start_from=0):
97102 all_images .extend ([
98103 np .squeeze (images [j ]).transpose ((1 , 2 , 0 )) for j in xrange (num_images )
99104 ])
100- labels = data ["labels" if cifar_version == "cifar10" else "fine_labels" ]
105+ labels = data [label_key ]
101106 all_labels .extend ([labels [j ] for j in xrange (num_images )])
102107 return image_utils .image_generator (
103108 all_images [start_from :start_from + how_many ],
@@ -417,3 +422,108 @@ def hparams(self, defaults, unused_model_hparams):
417422 p .max_expected_batch_size_per_shard = 4
418423 p .input_space_id = 1
419424 p .target_space_id = 1
425+
426+
427+ @registry .register_problem
428+ class ImageCifar20Tune (mnist .ImageMnistTune ):
429+ """Cifar-20 Tune."""
430+
431+ @property
432+ def num_classes (self ):
433+ return 20
434+
435+ @property
436+ def num_channels (self ):
437+ return 3
438+
439+ @property
440+ def class_labels (self ):
441+ return [
442+ "aquatic mammals" ,
443+ "fish" ,
444+ "flowers" ,
445+ "food containers" ,
446+ "fruit and vegetables" ,
447+ "household electrical devices" ,
448+ "household furniture" ,
449+ "insects" ,
450+ "large carnivores" ,
451+ "large man-made outdoor things" ,
452+ "large natural outdoor scenes" ,
453+ "large omnivores and herbivores" ,
454+ "medium-sized mammals" ,
455+ "non-insect invertebrates" ,
456+ "people" ,
457+ "reptiles" ,
458+ "small mammals" ,
459+ "trees" ,
460+ "vehicles 1" ,
461+ "vehicles 2" ,
462+ ]
463+
464+ def preprocess_example (self , example , mode , unused_hparams ):
465+ image = example ["inputs" ]
466+ image .set_shape ([_CIFAR100_IMAGE_SIZE , _CIFAR100_IMAGE_SIZE , 3 ])
467+ if mode == tf .estimator .ModeKeys .TRAIN :
468+ image = image_utils .cifar_image_augmentation (image )
469+ if not self ._was_reversed :
470+ image = tf .image .per_image_standardization (image )
471+ example ["inputs" ] = image
472+ return example
473+
474+ def generator (self , data_dir , tmp_dir , is_training ):
475+ if is_training :
476+ return cifar_generator ("cifar20" , tmp_dir , True , 48000 )
477+ else :
478+ return cifar_generator ("cifar20" , tmp_dir , True , 2000 , 48000 )
479+
480+
481+ @registry .register_problem
482+ class ImageCifar20 (ImageCifar20Tune ):
483+
484+ def generator (self , data_dir , tmp_dir , is_training ):
485+ if is_training :
486+ return cifar_generator ("cifar20" , tmp_dir , True , 50000 )
487+ else :
488+ return cifar_generator ("cifar20" , tmp_dir , False , 10000 )
489+
490+
491+ @registry .register_problem
492+ class ImageCifar20Plain (ImageCifar20 ):
493+
494+ def preprocess_example (self , example , mode , unused_hparams ):
495+ image = example ["inputs" ]
496+ image .set_shape ([_CIFAR100_IMAGE_SIZE , _CIFAR100_IMAGE_SIZE , 3 ])
497+ if not self ._was_reversed :
498+ image = tf .image .per_image_standardization (image )
499+ example ["inputs" ] = image
500+ return example
501+
502+
503+ @registry .register_problem
504+ class ImageCifar20PlainGen (ImageCifar20Plain ):
505+ """CIFAR-20 32x32 for image generation without standardization preprep."""
506+
507+ def dataset_filename (self ):
508+ return "image_cifar20_plain" # Reuse CIFAR-20 plain data.
509+
510+ def preprocess_example (self , example , mode , unused_hparams ):
511+ example ["inputs" ].set_shape ([_CIFAR100_IMAGE_SIZE , _CIFAR100_IMAGE_SIZE , 3 ])
512+ example ["inputs" ] = tf .to_int64 (example ["inputs" ])
513+ return example
514+
515+
516+ @registry .register_problem
517+ class ImageCifar20Plain8 (ImageCifar20 ):
518+ """CIFAR-20 rescaled to 8x8 for output: Conditional image generation."""
519+
520+ def dataset_filename (self ):
521+ return "image_cifar20_plain" # Reuse CIFAR-20 plain data.
522+
523+ def preprocess_example (self , example , mode , unused_hparams ):
524+ image = example ["inputs" ]
525+ image = image_utils .resize_by_area (image , 8 )
526+ if not self ._was_reversed :
527+ image = tf .image .per_image_standardization (image )
528+ example ["inputs" ] = image
529+ return example
0 commit comments