def prepare_image(self, im, buffers, params, key):
# im image shape is 500x300x3
params = dict(params)
params.update(self.params[key])
# change the color of the image
cvt = COLOR_CONVERSIONS[params.pop('colorspace', None)]
# change the image format and apply random augmentation to it given in the params
# Here we are not using any parameters as params has mostly nothing
# first get 318x318 image and then divide them in 106x106 tiles
randomize_image_shape = (576, 576)
image = cvt(
randomize_image(im,
randomize_image_shape,
background=flat(self.fill_values[key]),
**params))
# creating tiles from the image with shape 106x106
tiles = get_tile_images(image)
_, p, h, w, c = tiles.shape
tiles = tiles.reshape((p * p, h, w, c))
# applying random augmentations on the tiles
rng = rnd_parameters(None, None)
tiles = np.array([
randomize_image(tiles[i], (96, 96), flat(self.fill_values[key]),
True, **rng) for i in range(p * p)
])
# save some tiles for inspection
# for i in range(tiles.shape[0]):
# save_image(tiles[i], name = i+1)
tiles = np.array([hwc2chw(tiles[i]) for i in range(p * p)])
# here permute the tiles randomly and check if its working by saving them as png images
permute, index = random_permutation()
# probably write some list comprehension to permute first dim of tiles tensor
tiles = permutate_tiles(tiles, permute)
# generate ground truth for classifier
buffers['label'][...] = index
# Now tiles shape will be 9x3,96,96 and this will be used to create a batch
p, c, h, w = tiles.shape
tiles = (tiles.reshape((9 * c, h, w)))
# save the processed to the buffer with key
buffers[key][...] = tiles
buffers['target_image'][...] = tiles
# this is the processed image stored in buffer as numpy object of shape 3x224x224
return params
def prepare_image(self, im, buffers, params, key):
# im image shape is 500x300x3
params = dict(params)
params.update(self.params[key])
# change the color of the image
cvt = COLOR_CONVERSIONS[params.pop('colorspace', None)]
# change the image format and apply random augmentation to it given in the params
# Here we are not using any parameters as params has mostly nothing
# first get 318x318 image and then divide them in 106x106 tiles
randomize_image_shape = (576, 576)
image = cvt(randomize_image(
im, randomize_image_shape,
background=flat(self.fill_values[key]),
**params
))
# creating tiles from the image with shape 106x106
tiles = get_tile_images(image)
_, p, h, w, c = tiles.shape
tiles = tiles.reshape((p * p, h, w, c))
# applying random augmentations on the tiles
rng = rnd_parameters(None, None)
tiles = np.array([randomize_image(tiles[i], (96,96), flat(self.fill_values[key]), True, **rng) for i in range(p * p)])
# random crop tiles of 96x96x3
#tiles = crop_tiles(tiles)
tiles = np.array([hwc2chw(tiles[i]) for i in range(p * p)])
# Now tiles shape will be 9x3,96,96 and this will be used to create a batch
p, c, h, w = tiles.shape
tiles = (tiles.reshape((9 * c, h, w)))
# save the processed to the buffer with key
buffers[key][...] = tiles
buffers['target_image'][...] = tiles
# this is the processed image stored in buffer as numpy object of shape 3x224x224
return params