batch_size=1,

crop_size=0,

patch_size=256,

bbox= None,

augmentation=False):

'''

if np.ndim(mask) != 2 or np.ndim(image) != 3:

raise ValueError('image must have 3 dims and mask 2 dims')

if mask.shape != image.shape[1:]:

raise ValueError('image and mask shape is different')

im_max = float(np.max(image))

mask_max = 1.0

#select subimage

if bbox is not None:

# check bbox

if bbox[0] < 0 or bbox [2] < 0 \

or bbox[1] > mask.shape[0] or bbox[3] > mask.shape[0] \

or bbox[0] + patch_size > bbox[1] or bbox[2] + patch_size > bbox[3] \

or patch_size <= 0:

raise ValueError("Incorrect bbox or patch size")

img_ = image[:, bbox[0] : bbox[1], bbox[2]:bbox[3]]

mask_ = mask[bbox[0] : bbox[1], bbox[2]:bbox[3]]

else:

img_ = image

mask_ = mask

while 1:

x = []

y = []

for i in range (batch_size):

random_x = np.random.randint(0, mask_.shape[1] - patch_size)

random_y = np.random.randint(0, mask_.shape[0] - patch_size)

img_patch = img_[:,

random_y : random_y + patch_size,

random_x : random_x + patch_size] / im_max

# transform the image from channels-first (rasterio format) to channels-last (default tensorflow format)

img_patch = np.moveaxis(img_patch, 0, 2)

mask_patch = mask_[random_y : random_y + patch_size,

random_x : random_x + patch_size] / mask_max

if augmentation:

img_patch, mask_patch = random_augmentation(img_patch, mask_patch)

# mask is cropped as it may be useful for some convnets that have output size less than input

if crop_size > 0:

mask_patch = mask_patch[crop_size : -crop_size,

crop_size : -crop_size]

mask_patch = np.expand_dims(mask_patch, 2)

x.append(img_patch)

y.append(mask_patch)