def cifar10(mode, params):
"""CIFAR10 dataset creator."""
# Images are naturally 32x32.
model_mode = utils.estimator_mode_to_model_mode(mode)
hparams = params['hparams']
is_training = model_mode == enums.ModelMode.TRAIN
preprocessor = preprocessing.ImageToMultiViewedImagePreprocessor(
is_training=is_training,
preprocessing_options=hparams.input_data.preprocessing,
dataset_options=preprocessing.DatasetOptions(
decode_input=False,
image_mean_std=(np.array([[[-0.0172, -0.0356, -0.107]]]),
np.array([[[0.4046, 0.3988, 0.402]]]))),
bfloat16_supported=params['use_tpu'])
cifar_input = TfdsInput(
dataset_name='cifar10:3.*.*',
split='train' if is_training else 'test',
mode=model_mode,
preprocessor=preprocessor,
shard_per_host=hparams.input_data.shard_per_host,
cache=is_training,
shuffle_buffer=50000,
num_parallel_calls=64,
max_samples=hparams.input_data.max_samples,
label_noise_prob=hparams.input_data.label_noise_prob,
num_classes=get_num_classes(hparams),
data_dir=params['data_dir'],
)
return cifar_input.input_fn(params)
def imagenet(mode, params):
"""An input_fn for ImageNet (ILSVRC 2012) data."""
model_mode = utils.estimator_mode_to_model_mode(mode)
hparams = params['hparams']
is_training = model_mode == enums.ModelMode.TRAIN
preprocessor = preprocessing.ImageToMultiViewedImagePreprocessor(
is_training=is_training,
preprocessing_options=hparams.input_data.preprocessing,
dataset_options=preprocessing.DatasetOptions(decode_input=False),
bfloat16_supported=params['use_tpu'])
imagenet_input = TfdsInput(
dataset_name='imagenet2012:5.*.*',
split='train' if is_training else 'validation',
mode=model_mode,
preprocessor=preprocessor,
shuffle_buffer=1024,
shard_per_host=hparams.input_data.shard_per_host,
cache=is_training,
num_parallel_calls=64,
max_samples=hparams.input_data.max_samples,
label_noise_prob=hparams.input_data.label_noise_prob,
num_classes=get_num_classes(hparams),
data_dir=params['data_dir'],
)
return imagenet_input.input_fn(params)
def test_preprocess_image(self, decode_image, image_dtype, image_size,
is_training, augmentation_type, warp_prob,
augmentation_magnitude, eval_crop_method,
image_mean_std):
# tf.random.uniform() doesn't allow generating random values that are uint8.
image = tf.cast(
tf.random.uniform(
shape=(300, 400, 3), minval=0, maxval=255, dtype=tf.float32),
dtype=tf.uint8)
if decode_image:
image = tf.image.encode_jpeg(image)
else:
image = tf.cast(image, image_dtype)
expect_error = (not decode_image and image_dtype != tf.uint8)
if expect_error:
context_manager = self.assertRaises(AssertionError)
else:
context_manager = nullcontext()
with context_manager:
output = preprocessing.preprocess_image(
image,
is_training=is_training,
bfloat16_supported=False,
preprocessing_options=hparams.ImagePreprocessing(
image_size=image_size,
augmentation_type=augmentation_type,
warp_probability=warp_prob,
augmentation_magnitude=augmentation_magnitude,
eval_crop_method=eval_crop_method,
),
dataset_options=preprocessing.DatasetOptions(
image_mean_std=image_mean_std, decode_input=decode_image))
self.assertEqual(output.dtype, tf.float32)
self.assertEqual([image_size, image_size, 3], output.shape.as_list())
def test_input_class(self, input_class, model_mode, image_size,
max_samples):
split = 'train' if model_mode == enums.ModelMode.TRAIN else 'test'
batch_size = 2
dataset_size = 10
expected_num_batches = dataset_size // batch_size
if max_samples is not None and model_mode == enums.ModelMode.TRAIN:
expected_num_batches = max_samples // batch_size
params = {'batch_size': batch_size}
if input_class == 'TfdsInput':
with tfds.testing.mock_data(num_examples=dataset_size):
data = inputs.TfdsInput(
'cifar10',
split,
mode=model_mode,
preprocessor=preprocessing.
ImageToMultiViewedImagePreprocessor(
is_training=model_mode == enums.ModelMode.TRAIN,
preprocessing_options=hparams.ImagePreprocessing(
image_size=image_size, num_views=2),
dataset_options=preprocessing.DatasetOptions(
decode_input=False)),
max_samples=max_samples,
num_classes=10).input_fn(params)
else:
raise ValueError(f'Unknown input class {input_class}')
expected_num_channels = 3 if model_mode == enums.ModelMode.INFERENCE else 6
expected_batch_size = (None if model_mode == enums.ModelMode.INFERENCE
else batch_size)
if model_mode == enums.ModelMode.INFERENCE:
self.assertIsInstance(
data, tf.estimator.export.TensorServingInputReceiver)
image_shape = data.features.shape.as_list()
else:
self.assertIsInstance(data, tf.data.Dataset)
shapes = tf.data.get_output_shapes(data)
image_shape = shapes[0].as_list()
label_shape = shapes[1].as_list()
self.assertEqual([expected_batch_size], label_shape)
self.assertEqual([
expected_batch_size, image_size, image_size, expected_num_channels
], image_shape)
if model_mode == enums.ModelMode.INFERENCE:
return
# Now extract the Tensors
data = tf.data.make_one_shot_iterator(data).get_next()[0]
with self.cached_session() as sess:
for i in range(expected_num_batches + 1):
if i == expected_num_batches and model_mode == enums.ModelMode.EVAL:
with self.assertRaises(tf.errors.OutOfRangeError):
sess.run(data)
break
else:
sess.run(data)
def test_image_to_multi_viewed_image_preprocessor(self, is_training,
decode_image, num_views):
# tf.random.uniform() doesn't allow generating random values that are uint8.
image = tf.cast(
tf.random.uniform(
shape=(300, 400, 3), minval=0, maxval=255, dtype=tf.float32),
dtype=tf.uint8)
if decode_image:
image = tf.image.encode_jpeg(image)
image_size = 32
preprocessor = preprocessing.ImageToMultiViewedImagePreprocessor(
is_training=is_training,
preprocessing_options=hparams.ImagePreprocessing(
image_size=image_size, num_views=num_views),
dataset_options=preprocessing.DatasetOptions(decode_input=decode_image))
output = preprocessor.preprocess(image)
self.assertEqual(output.dtype, tf.float32)
self.assertEqual([image_size, image_size, 3 * num_views],
output.shape.as_list())
