def stack_images(stack_name, keys):
def stack(data):
images = [data[key] for key in keys]
data[stack_name] = tf.concat(images, 2)
return data
return preprocess.Preprocessor(stack)
def unstack_images(stack_name, keys):
def stack(data):
images = tf.unstack(data[stack_name], axis=2)
for key, image in zip(keys, images):
data[key] = image[:, :, None]
return data
return preprocess.Preprocessor(stack)
def augment_local_brightness(delta, seed):
def augment_local_brightness_(image):
# individual noise pattern
noise_pattern = make_noise_pattern(16, tf.shape(image), seed)
noise_factor = tf.truncated_normal((), 0.0, 0.5)
return tf.clip_by_value(image + delta * noise_pattern * noise_factor,
0.0, 1.0)
return preprocess.Preprocessor(augment_local_brightness_)
def augment_contrast(max_delta=0.0, lower=1.0, upper=1.0):
def augment_contrast_(image):
if max_delta > 0:
image = tf.image.random_brightness(image, max_delta)
if lower != 1.0 or upper != 1.0:
image = tf.image.random_contrast(image, lower, upper)
return image
return preprocess.Preprocessor(augment_contrast_)
def augment_local_contrast(contrast_factor, seed):
def augment_local_contrast_(image):
low_contrast_image = tf.image.adjust_contrast(image, contrast_factor)
# individual noise pattern
noise_pattern = make_noise_pattern(16, tf.shape(image), seed)
noise_factor = tf.truncated_normal((), 0.0, 0.5)
noise_pattern *= noise_factor
return image * (1.0 -
noise_pattern) + low_contrast_image * noise_pattern
return preprocess.Preprocessor(augment_local_contrast_)
def augment_with_noise(strength, seed=None):
def augment_with_noise_(image):
# individual noise pattern
noise_pattern = tf.random_uniform(tf.shape(image), -1, 1, seed=seed)
# a global noise strength factor
noise_strength = tf.truncated_normal((), 0.0, strength, seed=seed)
offset = noise_pattern * noise_strength
return tf.clip_by_value(image + offset, 0.0, 1.0)
return preprocess.Preprocessor(augment_with_noise_)
def downscale(mode="avg", factor=2):
@batched
def downscale_(image):
if mode == "avg":
return tf.layers.average_pooling2d(image,
factor,
factor,
padding="same")
elif mode == "max":
return tf.layers.max_pooling2d(image,
factor,
factor,
padding="same")
return preprocess.Preprocessor(downscale_)
def blur(length_scale, add_and_clip=False):
import scipy.ndimage
import numpy as np
def blur_fn(image):
return scipy.ndimage.gaussian_filter(image,
length_scale).astype(np.float32)
def blur_(image):
blurred = tf.py_func(blur_fn, [image], tf.float32, False)
blurred.set_shape(image.shape)
if add_and_clip:
blurred += image
blurred = tf.clip_by_value(blurred, 0.0, 1.0)
return blurred
return preprocess.Preprocessor(blur_)
def crop_and_resize_image(crop_size, image_size):
def crop_and_resize_image_(image):
channels = image.shape[2]
# cropping and resizing
if crop_size == "min":
cropping = tf.minimum(tf.shape(image)[0], tf.shape(image)[1])
elif crop_size == "max":
cropping = tf.maximum(tf.shape(image)[0], tf.shape(image)[1])
else:
cropping = crop_size
with tf.name_scope("crop_or_pad"):
cropped = tf.image.resize_image_with_crop_or_pad(
image, cropping, cropping)
tf.summary.image("cropped", cropped[None, :, :, :])
with tf.name_scope("resize"):
resized = tf.image.resize_images(cropped, [image_size, image_size])
resized.set_shape([image_size, image_size, channels])
tf.summary.image("resized", resized[None, :, :, :])
return tf.cast(resized, tf.uint8)
return preprocess.Preprocessor(crop_and_resize_image_)
def batched(f):
def g(arg):
return f(arg[None, ...])[0, ...]
return preprocess.Preprocessor(g)
def thicken():
def thicken_(image):
return tf.layers.max_pooling2d(tf.expand_dims(image, axis=0), 2, 1,
"same")[0, :, :, :]
return preprocess.Preprocessor(thicken_)