def get_real_activations(batch_size,
num_batches,
shuffle_buffer_size=100000,
split='validation',
get_logits=False):
"""Fetches batches inception pools and images.
NOTE: This function runs inference on an Inception network, so it would be
more efficient to run this on GPU or TPU than on CPU.
Args:
batch_size: The number of elements in a single minibatch.
num_batches: The number of batches to fetch at a time.
shuffle_buffer_size: The number of records to load before shuffling. Larger
means more likely randomization.
split: Shuffle if 'train', else deterministic.
get_logits: If `True`, return (logits, pools). Otherwise just return pools.
Returns:
A Tensor of `real_pools` or (`real_logits`, `real_pools`) with batch
dimension (batch_size * num_batches).
"""
ds = data_provider.provide_dataset(batch_size, shuffle_buffer_size, split)
ds = ds.map(lambda img, lbl: img) # Remove labels.
return get_activations_from_dataset(ds, num_batches, get_logits)
def train_eval_input_fn(mode, params):
"""Mode-aware input function."""
is_train = mode == tf.estimator.ModeKeys.TRAIN
split = 'train' if is_train else 'validation'
if params['tpu_params'].use_tpu_estimator:
bs = params['batch_size']
else:
bs = {
tf.estimator.ModeKeys.TRAIN: params['train_batch_size'],
tf.estimator.ModeKeys.EVAL: params['eval_batch_size'],
tf.estimator.ModeKeys.PREDICT: params['predict_batch_size'],
}[mode]
if params['debug_params'].fake_data:
fake_noise = tf.zeros([bs, params['z_dim']])
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.data.Dataset.from_tensors(fake_noise).repeat()
fake_imgs = tf.zeros([bs, 128, 128, 3])
fake_lbls = tf.zeros([bs], dtype=tf.int32)
ds = tf.data.Dataset.from_tensors((fake_noise, {
'images': fake_imgs,
'labels': fake_lbls
}))
ds = ds.repeat()
_verify_dataset_shape(ds, params['z_dim'])
return ds
num_towers = 1
def _make_noise(_):
noise = gen_module.make_z_normal(num_towers, bs, params['z_dim'])
return noise[0] # one tower
noise_ds = tf.data.Dataset.from_tensors(0).repeat().map(_make_noise)
if mode == tf.estimator.ModeKeys.PREDICT:
return noise_ds
images_ds = data_provider.provide_dataset(
bs, shuffle_buffer_size=params['shuffle_buffer_size'], split=split)
images_ds = images_ds.map(lambda img, lbl: {
'images': img,
'labels': lbl
}) # map to dict.
ds = tf.data.Dataset.zip((noise_ds, images_ds))
_verify_dataset_shape(ds, params['z_dim'])
return ds
def train_eval_input_fn(mode, params, restrict_classes=None, shift_classes=0):
"""Mode-aware input function.
restrict_classes: for use with intra fid
shift_classes: for use with restrict_classes
"""
is_train = mode == tf.estimator.ModeKeys.TRAIN
split = 'train' if is_train else flags.FLAGS.dataset_val_split_name
if params['tpu_params'].use_tpu_estimator:
bs = params['batch_size']
else:
bs = {
tf.estimator.ModeKeys.TRAIN: params['train_batch_size'],
tf.estimator.ModeKeys.EVAL: params['eval_batch_size'],
tf.estimator.ModeKeys.PREDICT: params['predict_batch_size'],
}[mode]
if params['debug_params'].fake_data:
fake_noise = tf.zeros([bs, params['z_dim']])
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.data.Dataset.from_tensors(fake_noise).repeat()
fake_imgs = tf.zeros(
[bs, flags.FLAGS.image_size, flags.FLAGS.image_size, 3])
fake_lbls = tf.zeros([bs], dtype=tf.int32)
ds = tf.data.Dataset.from_tensors((fake_noise, {
'images': fake_imgs,
'labels': fake_lbls
}))
ds = ds.repeat()
_verify_dataset_shape(ds, params['z_dim'])
return ds
num_towers = 1
def _make_noise(_):
noise = gen_module.make_z_normal(num_towers, bs, params['z_dim'])
return noise[0] # one tower
noise_ds = tf.data.Dataset.from_tensors(0).repeat().map(_make_noise)
if mode == tf.estimator.ModeKeys.PREDICT and not flags.FLAGS.mode == 'gen_images':
return noise_ds
images_ds = data_provider.provide_dataset(
bs,
shuffle_buffer_size=params['shuffle_buffer_size'],
split=split,
restrict_classes=restrict_classes)
if flags.FLAGS.unlabelled_dataset_name is not None:
unl_images_ds = data_provider_unlabelled.provide_dataset(
bs,
shuffle_buffer_size=params['shuffle_buffer_size'],
split=flags.FLAGS.unlabelled_dataset_split_name)
images_ds = tf.data.Dataset.zip((images_ds, unl_images_ds))
images_ds = images_ds.map(
lambda img_lab_tup, unl_img: {
'images': img_lab_tup[0],
'labels': img_lab_tup[1],
'unlabelled_images': unl_img
}) # map to dict.
else:
images_ds = images_ds.map(lambda img, lbl: {
'images': img,
'labels': lbl
}) # map to dict.
ds = tf.data.Dataset.zip((noise_ds, images_ds))
if restrict_classes is not None or flags.FLAGS.mode == 'intra_fid_eval':
ds = ds.map(lambda noise_ds, images_ds:
({
'z': noise_ds,
'labels': images_ds['labels'] - shift_classes
}, {
'images': images_ds['images'],
'labels': images_ds['labels'] - shift_classes
}))
elif flags.FLAGS.mode == 'gen_images':
def _make_labels_for_class(y):
return gen_module.make_one_batch_constant_labels(bs, y)
labels_todo = [900]
#labels_todo = list(range(flags.FLAGS.num_classes))
# hack to print your favorite classes
# labels_todo = list(sorted([130,96,90,88,164,175,281,289,290,292,294,323,441,475,555,581,607,654,661,663,688,779] * 5))
# okay, this is a ugly papercut, but an easy one-line modification of the above
def _make_one_batch_unif_random_labels(index):
del index
return gen_module.make_class_labels(bs, flags.FLAGS.num_classes)
if flags.FLAGS.gen_images_uniform_random_labels:
labs_ds = tf.data.Dataset.from_tensor_slices(
labels_todo).repeat().map(_make_one_batch_unif_random_labels)
else:
labs_ds = tf.data.Dataset.from_tensor_slices(
labels_todo).repeat().map(_make_labels_for_class)
ds = tf.data.Dataset.zip((noise_ds, images_ds, labs_ds))
ds = ds.map(lambda noise_ds_, images_ds_, labs_ds_: {
'z': noise_ds_,
'labels': labs_ds_
}) # fake data only
# ds = ds.map(lambda noise_ds_, images_ds_, labs_ds_: ({'z': noise_ds_, 'labels': labs_ds_}, images_ds_) )
else:
_verify_dataset_shape(ds, params['z_dim'])
return ds