def _get_batches_of_transformed_samples(self, index_array):
V_red = self.image_data_generator.V_red
batch_x = np.zeros(
(len(index_array), V_red.shape[1]), dtype=self.dtype
) #tuple([len(index_array)] + list(self.x.shape)[1:]),
# dtype=self.dtype)
for i, j in enumerate(index_array):
x = self.x[j]
params = self.image_data_generator.get_random_transform(x.shape)
x = self.image_data_generator.apply_transform(
x.astype(self.dtype), params)
# x = self.image_data_generator.standardize(x)
batch_x[i] = np.dot(np.reshape(x, (1, 2916)), V_red)
if self.save_to_dir:
for i, j in enumerate(index_array):
img = array_to_img(batch_x[i], self.data_format, scale=True)
fname = '{prefix}_{index}_{hash}.{format}'.format(
prefix=self.save_prefix,
index=j,
hash=np.random.randint(1e4),
format=self.save_format)
img.save(os.path.join(self.save_to_dir, fname))
batch_x_miscs = [xx[index_array] for xx in self.x_misc]
output = (batch_x if batch_x_miscs == [] else [batch_x] +
batch_x_miscs, )
if self.y is None:
return output[0]
output += (self.y[index_array], )
if self.sample_weight is not None:
output += (self.sample_weight[index_array], )
return output
def _get_batches_of_transformed_samples(self, index_array):
# build batch of image data
batch_x = np.array([self.x[j] for j in index_array])
# transform the image data
batch_x = np.array(
[self.image_data_generator.transform_image(x) for x in batch_x])
if self.y is not None:
batch_y = np.array([self.y[j] for j in index_array])
else:
batch_y = np.array([])
if self.save_to_dir:
for i, j in enumerate(index_array):
img = array_to_img(batch_x[i], self.data_format, scale=True)
fname = '{prefix}_{index}_{hash}.{format}'.format(
prefix=self.save_prefix,
index=j,
hash=np.random.randint(1e4),
format=self.save_format)
img.save(os.path.join(self.save_to_dir, fname))
batch_x_miscs = [xx[index_array] for xx in self.x_misc]
output = (batch_x if batch_x_miscs == [] else [batch_x] +
batch_x_miscs, )
if self.y is None:
return output[0]
output += (batch_y, )
if self.sample_weight is not None:
output += (self.sample_weight[index_array], )
return output
def _get_batches_of_transformed_samples(self, index_array):
"""Gets a batch of transformed samples.
# Arguments
index_array: Array of sample indices to include in batch.
# Returns
A batch of transformed samples.
"""
# build batch of image data
# self.filepaths is dynamic, is better to call it once outside the loop
filepaths = self.filepaths
# build batch of image data
batch_x = np.array([
load_img(filepaths[x],
color_mode=self.color_mode,
target_size=self.target_size,
interpolation=self.interpolation) for x in index_array
])
# transform the image data
batch_x = np.array(
[self.image_data_generator.transform_image(x) for x in batch_x])
# optionally save augmented images to disk for debugging purposes
if self.save_to_dir:
for i, j in enumerate(index_array):
img = array_to_img(batch_x[i], self.data_format, scale=True)
fname = '{prefix}_{index}_{hash}.{format}'.format(
prefix=self.save_prefix,
index=j,
hash=np.random.randint(1e7),
format=self.save_format)
img.save(os.path.join(self.save_to_dir, fname))
# build batch of labels
if self.class_mode == 'input':
batch_y = batch_x.copy()
elif self.class_mode in {'binary', 'sparse'}:
batch_y = np.empty(len(batch_x), dtype=self.dtype)
for i, n_observation in enumerate(index_array):
batch_y[i] = self.classes[n_observation]
elif self.class_mode == 'categorical':
batch_y = np.zeros((len(batch_x), len(self.class_indices)),
dtype=self.dtype)
for i, n_observation in enumerate(index_array):
batch_y[i, self.classes[n_observation]] = 1.
elif self.class_mode == 'multi_output':
batch_y = [output[index_array] for output in self.labels]
elif self.class_mode == 'raw':
batch_y = self.labels[index_array]
else:
return batch_x
if self.sample_weight is None:
return batch_x, batch_y
else:
return batch_x, batch_y, self.sample_weight[index_array]
def test_array_to_img_and_img_to_array():
height, width = 10, 8
# Test the data format
# Test RGB 3D
x = np.random.random((3, height, width))
img = utils.array_to_img(x, data_format='channels_first')
assert img.size == (width, height)
x = utils.img_to_array(img, data_format='channels_first')
assert x.shape == (3, height, width)
# Test RGBA 3D
x = np.random.random((4, height, width))
img = utils.array_to_img(x, data_format='channels_first')
assert img.size == (width, height)
x = utils.img_to_array(img, data_format='channels_first')
assert x.shape == (4, height, width)
# Test 2D
x = np.random.random((1, height, width))
img = utils.array_to_img(x, data_format='channels_first')
assert img.size == (width, height)
x = utils.img_to_array(img, data_format='channels_first')
assert x.shape == (1, height, width)
# grayscale 32-bit signed integer
x = np.array(
np.random.randint(-2147483648, 2147483647, (1, height, width)),
dtype=np.int32
)
img = utils.array_to_img(x, data_format='channels_first')
assert img.size == (width, height)
x = utils.img_to_array(img, data_format='channels_first')
assert x.shape == (1, height, width)
# Test tf data format
# Test RGB 3D
x = np.random.random((height, width, 3))
img = utils.array_to_img(x, data_format='channels_last')
assert img.size == (width, height)
x = utils.img_to_array(img, data_format='channels_last')
assert x.shape == (height, width, 3)
# Test RGBA 3D
x = np.random.random((height, width, 4))
img = utils.array_to_img(x, data_format='channels_last')
assert img.size == (width, height)
x = utils.img_to_array(img, data_format='channels_last')
assert x.shape == (height, width, 4)
# Test 2D
x = np.random.random((height, width, 1))
img = utils.array_to_img(x, data_format='channels_last')
assert img.size == (width, height)
x = utils.img_to_array(img, data_format='channels_last')
assert x.shape == (height, width, 1)
# grayscale 16-bit signed integer
x = np.array(
np.random.randint(-2147483648, 2147483647, (height, width, 1)),
dtype=np.int16
)
img = utils.array_to_img(x, data_format='channels_last')
assert img.size == (width, height)
x = utils.img_to_array(img, data_format='channels_last')
assert x.shape == (height, width, 1)
# grayscale 32-bit signed integer
x = np.array(
np.random.randint(-2147483648, 2147483647, (height, width, 1)),
dtype=np.int32
)
img = utils.array_to_img(x, data_format='channels_last')
assert img.size == (width, height)
x = utils.img_to_array(img, data_format='channels_last')
assert x.shape == (height, width, 1)
# Test invalid use case
with pytest.raises(ValueError):
x = np.random.random((height, width)) # not 3D
img = utils.array_to_img(x, data_format='channels_first')
with pytest.raises(ValueError):
x = np.random.random((height, width, 3))
# unknown data_format
img = utils.array_to_img(x, data_format='channels')
with pytest.raises(ValueError):
# neither RGB, RGBA, or gray-scale
x = np.random.random((height, width, 5))
img = utils.array_to_img(x, data_format='channels_last')
with pytest.raises(ValueError):
x = np.random.random((height, width, 3))
# unknown data_format
img = utils.img_to_array(x, data_format='channels')
with pytest.raises(ValueError):
# neither RGB, RGBA, or gray-scale
x = np.random.random((height, width, 5, 3))
img = utils.img_to_array(x, data_format='channels_last')
def test_load_img(tmpdir):
filename_rgb = str(tmpdir / 'rgb_utils.png')
filename_rgba = str(tmpdir / 'rgba_utils.png')
filename_grayscale_8bit = str(tmpdir / 'grayscale_8bit_utils.png')
filename_grayscale_16bit = str(tmpdir / 'grayscale_16bit_utils.tiff')
filename_grayscale_32bit = str(tmpdir / 'grayscale_32bit_utils.tiff')
original_rgb_array = np.array(255 * np.random.rand(100, 100, 3),
dtype=np.uint8)
original_rgb = utils.array_to_img(original_rgb_array, scale=False)
original_rgb.save(filename_rgb)
original_rgba_array = np.array(255 * np.random.rand(100, 100, 4),
dtype=np.uint8)
original_rgba = utils.array_to_img(original_rgba_array, scale=False)
original_rgba.save(filename_rgba)
original_grayscale_8bit_array = np.array(255 * np.random.rand(100, 100, 1),
dtype=np.uint8)
original_grayscale_8bit = utils.array_to_img(original_grayscale_8bit_array,
scale=False)
original_grayscale_8bit.save(filename_grayscale_8bit)
original_grayscale_16bit_array = np.array(
np.random.randint(-2147483648, 2147483647, (100, 100, 1)), dtype=np.int16
)
original_grayscale_16bit = utils.array_to_img(original_grayscale_16bit_array,
scale=False, dtype='int16')
original_grayscale_16bit.save(filename_grayscale_16bit)
original_grayscale_32bit_array = np.array(
np.random.randint(-2147483648, 2147483647, (100, 100, 1)), dtype=np.int32
)
original_grayscale_32bit = utils.array_to_img(original_grayscale_32bit_array,
scale=False, dtype='int32')
original_grayscale_32bit.save(filename_grayscale_32bit)
# Test that loaded image is exactly equal to original.
loaded_im = utils.load_img(filename_rgb)
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_rgb_array.shape
assert np.all(loaded_im_array == original_rgb_array)
loaded_im = utils.load_img(filename_rgba, color_mode='rgba')
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_rgba_array.shape
assert np.all(loaded_im_array == original_rgba_array)
loaded_im = utils.load_img(filename_rgb, color_mode='grayscale')
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (original_rgb_array.shape[0],
original_rgb_array.shape[1], 1)
loaded_im = utils.load_img(filename_grayscale_8bit, color_mode='grayscale')
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_grayscale_8bit_array.shape
assert np.all(loaded_im_array == original_grayscale_8bit_array)
loaded_im = utils.load_img(filename_grayscale_16bit, color_mode='grayscale')
loaded_im_array = utils.img_to_array(loaded_im, dtype='int16')
assert loaded_im_array.shape == original_grayscale_16bit_array.shape
assert np.all(loaded_im_array == original_grayscale_16bit_array)
# test casting int16 image to float32
loaded_im_array = utils.img_to_array(loaded_im)
assert np.allclose(loaded_im_array, original_grayscale_16bit_array)
loaded_im = utils.load_img(filename_grayscale_32bit, color_mode='grayscale')
loaded_im_array = utils.img_to_array(loaded_im, dtype='int32')
assert loaded_im_array.shape == original_grayscale_32bit_array.shape
assert np.all(loaded_im_array == original_grayscale_32bit_array)
# test casting int32 image to float32
loaded_im_array = utils.img_to_array(loaded_im)
assert np.allclose(loaded_im_array, original_grayscale_32bit_array)
# Test that nothing is changed when target size is equal to original.
loaded_im = utils.load_img(filename_rgb, target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_rgb_array.shape
assert np.all(loaded_im_array == original_rgb_array)
loaded_im = utils.load_img(filename_rgba, color_mode='rgba',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_rgba_array.shape
assert np.all(loaded_im_array == original_rgba_array)
loaded_im = utils.load_img(filename_rgb, color_mode='grayscale',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (original_rgba_array.shape[0],
original_rgba_array.shape[1], 1)
loaded_im = utils.load_img(filename_grayscale_8bit, color_mode='grayscale',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_grayscale_8bit_array.shape
assert np.all(loaded_im_array == original_grayscale_8bit_array)
loaded_im = utils.load_img(filename_grayscale_16bit, color_mode='grayscale',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im, dtype='int16')
assert loaded_im_array.shape == original_grayscale_16bit_array.shape
assert np.all(loaded_im_array == original_grayscale_16bit_array)
loaded_im = utils.load_img(filename_grayscale_32bit, color_mode='grayscale',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im, dtype='int32')
assert loaded_im_array.shape == original_grayscale_32bit_array.shape
assert np.all(loaded_im_array == original_grayscale_32bit_array)
# Test down-sampling with bilinear interpolation.
loaded_im = utils.load_img(filename_rgb, target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 3)
loaded_im = utils.load_img(filename_rgba, color_mode='rgba',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 4)
loaded_im = utils.load_img(filename_rgb, color_mode='grayscale',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_8bit, color_mode='grayscale',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_16bit, color_mode='grayscale',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im, dtype='int16')
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_32bit, color_mode='grayscale',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im, dtype='int32')
assert loaded_im_array.shape == (25, 25, 1)
# Test down-sampling with nearest neighbor interpolation.
loaded_im_nearest = utils.load_img(filename_rgb, target_size=(25, 25),
interpolation="nearest")
loaded_im_array_nearest = utils.img_to_array(loaded_im_nearest)
assert loaded_im_array_nearest.shape == (25, 25, 3)
assert np.any(loaded_im_array_nearest != loaded_im_array)
loaded_im_nearest = utils.load_img(filename_rgba, color_mode='rgba',
target_size=(25, 25),
interpolation="nearest")
loaded_im_array_nearest = utils.img_to_array(loaded_im_nearest)
assert loaded_im_array_nearest.shape == (25, 25, 4)
assert np.any(loaded_im_array_nearest != loaded_im_array)
loaded_im = utils.load_img(filename_grayscale_8bit, color_mode='grayscale',
target_size=(25, 25), interpolation="nearest")
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_16bit, color_mode='grayscale',
target_size=(25, 25), interpolation="nearest")
loaded_im_array = utils.img_to_array(loaded_im, dtype='int16')
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_32bit, color_mode='grayscale',
target_size=(25, 25), interpolation="nearest")
loaded_im_array = utils.img_to_array(loaded_im, dtype='int32')
assert loaded_im_array.shape == (25, 25, 1)
# Check that exception is raised if interpolation not supported.
loaded_im = utils.load_img(filename_rgb, interpolation="unsupported")
with pytest.raises(ValueError):
loaded_im = utils.load_img(filename_rgb, target_size=(25, 25),
interpolation="unsupported")
def _get_batches_of_transformed_samples(self, index_array):
"""Gets a batch of transformed samples.
# Arguments
index_array: Array of sample indices to include in batch.
# Returns
A batch of transformed samples in one-hot-encoded format.
"""
additional_bg_class = 0
if self.include_background:
additional_bg_class = 1
batch_x = np.zeros(
(len(index_array), ) + (self.target_size[0], self.target_size[1],
self.num_classes + additional_bg_class),
dtype=self.dtype)
# build batch of image data
# self.filepaths is dynamic, is better to call it once outside the loop
filepaths = self.filepaths
known_label_keys = self.dropped_labels_memory.keys()
known_label_drops = {}
unset_label_drops = []
for i, j in enumerate(index_array):
if j in known_label_keys:
known_label_drops[i] = j
else:
unset_label_drops.append(i)
one_hot_map = np.zeros((self.target_size[0], self.target_size[1],
self.num_classes + additional_bg_class),
dtype=np.float32)
# Iterate over all classes
params = None
reserved_pixels = None
for k in range(self.num_classes):
filepath = os.path.join(self.directory, self._mask_classes[k],
filepaths[j])
img = load_img(filepath,
color_mode=self.color_mode,
target_size=self.target_size,
interpolation=self.interpolation)
x = img_to_array(img, data_format=self.data_format)
# Pillow images should be closed after `load_img`,
# but not PIL images.
if hasattr(img, 'close'):
img.close()
if self.image_data_generator:
# Params need to be set once for every image (not for every mask)
if params is None:
params = self.image_data_generator.get_random_transform(
x.shape)
x = self.image_data_generator.apply_transform(x, params)
x = self.image_data_generator.standardize(x)
if reserved_pixels is None:
reserved_pixels = np.zeros(x.shape)
x = np.where(reserved_pixels, 0, x)
reserved_pixels += x
one_hot_map += self.get_one_hot_map(
x,
k,
background=self.background_color,
additional_bg_class=additional_bg_class)
# If one_hot_map has a max value >1 whe have overlapping classes -> prohibited
one_hot_map = one_hot_map.numpy()
if one_hot_map.max() > 1:
raise ValueError(
'Mask mismatch: classes are not mutually exclusive (multiple class definitions for '
'one pixel).')
if self.include_background:
# Background class is everywhere, where the one-hot encoding has only zeros
one_hot_map = tf.where(
tf.repeat(tf.reshape(
tf.math.count_nonzero(
one_hot_map == tf.zeros(self.num_classes + 1),
axis=2), [img.height, img.width, 1]),
self.num_classes + 1,
axis=2) == self.num_classes + 1,
tf.one_hot(
tf.constant(self.num_classes,
shape=one_hot_map.shape[:2]),
self.num_classes + 1), one_hot_map)
batch_x[i] = one_hot_map
# optionally save augmented images to disk for debugging purposes
if self.save_to_dir:
for i, j in enumerate(index_array):
img = array_to_img(batch_x[i], self.data_format, scale=True)
fname = '{prefix}_{index}_{hash}.{format}'.format(
prefix=self.save_prefix,
index=j,
hash=np.random.randint(1e7),
format=self.save_format)
img.save(os.path.join(self.save_to_dir, fname))
# Only where labels are not already known
if len(self.heterogeneously_labeled_masks) > 0:
batch_x[unset_label_drops] = self.remove_masks(
np.take(batch_x, unset_label_drops, axis=0))
# Known labels
for item in known_label_drops.items():
index_in_batch = item[0]
index_in_memory = item[1]
memory_binary_mask = self.dropped_labels_memory.get(
index_in_memory)
batch_x[index_in_batch, :, :, :][
memory_binary_mask] = DELETED_MASK_IDENTIFIER
# Extend Memory of known deletion masks.
delete_mask = np.where(batch_x == DELETED_MASK_IDENTIFIER, True,
False)
for i in range(len(index_array)):
self.dropped_labels_memory[index_array[i]] = delete_mask[
i, :, :, :]
return batch_x
def _get_batches_of_transformed_samples(self, index_array):
# IMPORTANT: the next line is changed with respect to the original
# keras implementation
# Change: self.target_shape <-- list(self.x.shape)[1:]
# Additionally: support for more than one augmentation per image
batch_x = np.zeros(tuple([len(index_array) * self.n_aug] \
+ self.target_shape), dtype=self.dtype)
batch_y_id = np.zeros([len(index_array) * self.n_aug] * 2,
dtype=np.uint8)
for i, j in enumerate(index_array):
batch_y_id[i * self.n_aug:i * self.n_aug + self.n_aug,
i * self.n_aug:i * self.n_aug + self.n_aug] = 1
for k in range(self.n_aug):
x = self.x[j]
params = self.image_data_generator.get_random_transform(
x.shape)
x = self.image_data_generator.apply_transform(
x.astype(self.dtype), params)
x = self.image_data_generator.standardize(x)
batch_x[i * self.n_aug + k] = x
if self.n_aug > 1:
batch_y_id[np.tril_indices(batch_y_id.shape[0])] = 0
if self.save_to_dir:
for i, j in enumerate(index_array):
img = array_to_img(batch_x[i], self.data_format, scale=True)
fname = '{prefix}_{index}_{hash}.{format}'.format(
prefix=self.save_prefix,
index=j,
hash=np.random.randint(1e4),
format=self.save_format)
img.save(os.path.join(self.save_to_dir, fname))
# Re-shuffle, in order to avoid contiguous augmented images
if (self.n_aug > 1) & (self.shuffle):
if self.seed is not None:
np.random.seed(self.seed + self.total_batches_seen)
idx = np.random.permutation(batch_x.shape[0])
else:
idx = np.arange(batch_x.shape[0])
batch_x = batch_x[idx]
batch_x_miscs = [xx[index_array][idx] for xx in self.x_misc]
output = (batch_x if batch_x_miscs == [] else [batch_x] +
batch_x_miscs, )
if self.y is None:
return output[0]
batch_y_cat = np.repeat(self.y[index_array], self.n_aug, axis=0)[idx]
batch_y_id = batch_y_id[idx][:, idx]
batch_y = [batch_y_cat, batch_y_id]
batch_y_miscs = [yy[indey_array][idx] for yy in self.y_misc]
output += (batch_y if batch_y_miscs == [] else batch_y +
batch_y_miscs, )
if self.sample_weight is not None:
output += (np.repeat(self.sample_weight[index_array],
self.n_aug,
axis=0)[idx], )
return output
def _get_batches_of_transformed_samples(self, index_array):
"""Gets a batch of transformed samples.
# Arguments
index_array: Array of sample indices to include in batch.
# Returns
A batch of transformed samples.
"""
index_array = self.X[index_array]
batch_x = np.zeros((len(index_array), ) + self.image_shape,
dtype=self.dtype)
# build batch of image data
# self.filepaths is dynamic, is better to call it once outside the loop
filepaths = self.filepaths
for i, j in enumerate(index_array):
img = load_img(filepaths[j],
color_mode=self.color_mode,
target_size=self.target_size,
interpolation=self.interpolation)
x = img_to_array(img, data_format=self.data_format)
# Pillow images should be closed after `load_img`,
# but not PIL images.
if hasattr(img, 'close'):
img.close()
if self.image_data_generator:
params = self.image_data_generator.get_random_transform(
x.shape)
x = self.image_data_generator.apply_transform(x, params)
x = self.image_data_generator.standardize(x)
batch_x[i] = x
# optionally save augmented images to disk for debugging purposes
if self.save_to_dir:
for i, j in enumerate(index_array):
img = array_to_img(batch_x[i], self.data_format, scale=True)
fname = '{prefix}_{index}_{hash}.{format}'.format(
prefix=self.save_prefix,
index=j,
hash=np.random.randint(1e7),
format=self.save_format)
img.save(os.path.join(self.save_to_dir, fname))
# build batch of labels
if self.class_mode == 'input':
batch_y = batch_x.copy()
elif self.class_mode in {'binary', 'sparse'}:
batch_y = np.empty(len(batch_x), dtype=self.dtype)
for i, n_observation in enumerate(index_array):
batch_y[i] = self.classes[n_observation]
elif self.class_mode == 'categorical':
batch_y = np.zeros((len(batch_x), len(self.class_indices)),
dtype=self.dtype)
for i, n_observation in enumerate(index_array):
batch_y[i, self.classes[n_observation]] = 1.
elif self.class_mode == 'multi_output':
batch_y = [output[index_array] for output in self.labels]
elif self.class_mode == 'raw':
batch_y = self.labels[index_array]
else:
return batch_x
if self.sample_weight is None:
return batch_x, batch_y
else:
return batch_x, batch_y, self.sample_weight[index_array]
def sample(self, X=None, sample_size=None, standardize=True):
""" Retrived fix-sized image tersors
Parameters
----------
X : 2D-array. Default is None
Expanded sub-index array of the dataframe.
If None, X = np.arange(n_samples)[:, np.newaxis].
sample_size : int. Default is None.
The number of samples to be retrieved.
If None, sample_size = X.shape[0]
standardize : bool. Default is True.
Whether to transform the image tersor data.
If False, return direct results of `img_to_array`.
"""
if X is None:
X = np.arange(self.dataframe.shape[0])[:, np.newaxis]
if not sample_size:
sample_size = X.shape[0]
retrieved_X = np.zeros((sample_size, ) + self.image_shape,
dtype=self.dtype)
filepaths = self.filepaths
indices = np.squeeze(X)
sample_index = self.rng_.choice(indices,
size=sample_size,
replace=False)
for i, j in enumerate(sample_index):
img = load_img(filepaths[j],
color_mode=self.color_mode,
target_size=self.target_size,
interpolation=self.interpolation)
x = img_to_array(img, data_format=self.data_format)
if hasattr(img, 'close'):
img.close()
if not standardize:
retrieved_X[i] = x
continue
params = self.get_random_transform(x.shape)
x = self.apply_transform(x, params)
x = self.standardize(x)
retrieved_X[i] = x
if self.save_to_dir:
for i, j in enumerate(sample_index):
img = array_to_img(retrieved_X[i],
self.data_format,
scale=True)
fname = '{prefix}_{index}_{hash}.{format}'.format(
prefix=self.save_prefix,
index=j,
hash=np.random.randint(1e7),
format=self.save_format)
img.save(os.path.join(self.save_to_dir, fname))
# retrieve labels
if self.class_mode == 'input':
retrieved_y = retrieved_X.copy()
elif self.class_mode in {'binary', 'sparse'}:
retrieved_y = np.empty(sample_size, dtype=self.dtype)
for i, n_observation in enumerate(sample_index):
retrieved_y[i] = self.classes[n_observation]
elif self.class_mode == 'categorical':
retrieved_y = np.zeros((sample_size, len(self.class_indices)),
dtype=self.dtype)
for i, n_observation in enumerate(sample_index):
retrieved_y[i, self.classes[n_observation]] = 1.
elif self.class_mode == 'multi_output':
retrieved_y = [output[sample_index] for output in self.labels]
elif self.class_mode == 'raw':
retrieved_y = self.labels[sample_index]
else:
return retrieved_X
return retrieved_X, retrieved_y
def _get_batches_of_transformed_samples(self, index_array):
"""Gets a batch of transformed samples.
# Arguments
index_array: Array of sample indices to include in batch.
# Returns
A batch of transformed samples.
"""
y = self.classes[index_array]
# creo una matrice indice - classe
batch = np.concatenate(
[np.expand_dims(index_array, 1),
np.expand_dims(y, 1)], axis=1)
batch = pd.DataFrame(batch, columns=['idx', 'class'])
batch = batch.groupby('class')
# Per ogni classe genero K sample
# batch totale 128
try:
batch = batch.apply(
lambda _x: _x.sample(18).reset_index(drop=True))
except ValueError:
batch = batch.apply(
lambda _x: _x.sample(18, replace=True).reset_index(drop=True))
print("This batch is garbage")
index_array = np.array(batch['idx'])
index_array = np.random.permutation(index_array)
batch_x = np.zeros((len(index_array), ) + self.image_shape,
dtype=self.dtype)
# build batch of image data
# self.filepaths is dynamic, is better to call it once outside the loop
filepaths = self.filepaths
for i, j in enumerate(index_array):
img = load_img(filepaths[j],
color_mode=self.color_mode,
target_size=self.target_size,
interpolation=self.interpolation)
x = img_to_array(img, data_format=self.data_format)
# Pillow images should be closed after `load_img`,
# but not PIL images.
if hasattr(img, 'close'):
img.close()
if self.image_data_generator:
params = self.image_data_generator.get_random_transform(
x.shape)
x = self.image_data_generator.apply_transform(x, params)
x = self.image_data_generator.standardize(x)
batch_x[i] = x
# optionally save augmented images to disk for debugging purposes
if self.save_to_dir:
for i, j in enumerate(index_array):
img = array_to_img(batch_x[i], self.data_format, scale=True)
fname = '{prefix}_{index}_{hash}.{format}'.format(
prefix=self.save_prefix,
index=j,
hash=np.random.randint(1e7),
format=self.save_format)
img.save(os.path.join(self.save_to_dir, fname))
# build batch of labels
if self.class_mode == 'input':
batch_y = batch_x.copy()
elif self.class_mode in {'binary', 'sparse'}:
batch_y = np.empty(len(batch_x), dtype=self.dtype)
for i, n_observation in enumerate(index_array):
batch_y[i] = self.classes[n_observation]
elif self.class_mode == 'categorical':
batch_y = np.zeros((len(batch_x), len(self.class_indices)),
dtype=self.dtype)
for i, n_observation in enumerate(index_array):
batch_y[i, self.classes[n_observation]] = 1.
elif self.class_mode == 'multi_output':
batch_y = [output[index_array] for output in self.labels]
elif self.class_mode == 'raw':
batch_y = self.labels[index_array]
else:
return batch_x
if self.sample_weight is None:
return [batch_x, batch_x, batch_x, np.argmax(batch_y, axis=1)],\
[np.zeros((len(batch_x), 4096), dtype='float16'), batch_y]
else:
return batch_x, batch_y, self.sample_weight[index_array]
def write_sample_image(tmpdir):
im = utils.array_to_img(np.random.rand(1, 1, 3))
path = str(tmpdir / 'sample_image.png')
utils.save_img(path, im)
return path
def test_load_img(tmpdir):
filename_rgb = str(tmpdir / 'rgb_utils.png')
filename_rgba = str(tmpdir / 'rgba_utils.png')
filename_grayscale_8bit = str(tmpdir / 'grayscale_8bit_utils.png')
filename_grayscale_16bit = str(tmpdir / 'grayscale_16bit_utils.tiff')
filename_grayscale_32bit = str(tmpdir / 'grayscale_32bit_utils.tiff')
original_rgb_array = np.array(255 * np.random.rand(100, 100, 3),
dtype=np.uint8)
original_rgb = utils.array_to_img(original_rgb_array, scale=False)
original_rgb.save(filename_rgb)
original_rgba_array = np.array(255 * np.random.rand(100, 100, 4),
dtype=np.uint8)
original_rgba = utils.array_to_img(original_rgba_array, scale=False)
original_rgba.save(filename_rgba)
original_grayscale_8bit_array = np.array(255 * np.random.rand(100, 100, 1),
dtype=np.uint8)
original_grayscale_8bit = utils.array_to_img(original_grayscale_8bit_array,
scale=False)
original_grayscale_8bit.save(filename_grayscale_8bit)
original_grayscale_16bit_array = np.array(np.random.randint(
-2147483648, 2147483647, (100, 100, 1)),
dtype=np.int16)
original_grayscale_16bit = utils.array_to_img(
original_grayscale_16bit_array, scale=False, dtype='int16')
original_grayscale_16bit.save(filename_grayscale_16bit)
original_grayscale_32bit_array = np.array(np.random.randint(
-2147483648, 2147483647, (100, 100, 1)),
dtype=np.int32)
original_grayscale_32bit = utils.array_to_img(
original_grayscale_32bit_array, scale=False, dtype='int32')
original_grayscale_32bit.save(filename_grayscale_32bit)
# Test that loaded image is exactly equal to original.
loaded_im = utils.load_img(filename_rgb)
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_rgb_array.shape
assert np.all(loaded_im_array == original_rgb_array)
loaded_im = utils.load_img(filename_rgba, color_mode='rgba')
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_rgba_array.shape
assert np.all(loaded_im_array == original_rgba_array)
loaded_im = utils.load_img(filename_rgb, color_mode='grayscale')
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (original_rgb_array.shape[0],
original_rgb_array.shape[1], 1)
loaded_im = utils.load_img(filename_grayscale_8bit, color_mode='grayscale')
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_grayscale_8bit_array.shape
assert np.all(loaded_im_array == original_grayscale_8bit_array)
loaded_im = utils.load_img(filename_grayscale_16bit,
color_mode='grayscale')
loaded_im_array = utils.img_to_array(loaded_im, dtype='int16')
assert loaded_im_array.shape == original_grayscale_16bit_array.shape
assert np.all(loaded_im_array == original_grayscale_16bit_array)
# test casting int16 image to float32
loaded_im_array = utils.img_to_array(loaded_im)
assert np.allclose(loaded_im_array, original_grayscale_16bit_array)
loaded_im = utils.load_img(filename_grayscale_32bit,
color_mode='grayscale')
loaded_im_array = utils.img_to_array(loaded_im, dtype='int32')
assert loaded_im_array.shape == original_grayscale_32bit_array.shape
assert np.all(loaded_im_array == original_grayscale_32bit_array)
# test casting int32 image to float32
loaded_im_array = utils.img_to_array(loaded_im)
assert np.allclose(loaded_im_array, original_grayscale_32bit_array)
# Test that nothing is changed when target size is equal to original.
loaded_im = utils.load_img(filename_rgb, target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_rgb_array.shape
assert np.all(loaded_im_array == original_rgb_array)
loaded_im = utils.load_img(filename_rgba,
color_mode='rgba',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_rgba_array.shape
assert np.all(loaded_im_array == original_rgba_array)
loaded_im = utils.load_img(filename_rgb,
color_mode='grayscale',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (original_rgba_array.shape[0],
original_rgba_array.shape[1], 1)
loaded_im = utils.load_img(filename_grayscale_8bit,
color_mode='grayscale',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == original_grayscale_8bit_array.shape
assert np.all(loaded_im_array == original_grayscale_8bit_array)
loaded_im = utils.load_img(filename_grayscale_16bit,
color_mode='grayscale',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im, dtype='int16')
assert loaded_im_array.shape == original_grayscale_16bit_array.shape
assert np.all(loaded_im_array == original_grayscale_16bit_array)
loaded_im = utils.load_img(filename_grayscale_32bit,
color_mode='grayscale',
target_size=(100, 100))
loaded_im_array = utils.img_to_array(loaded_im, dtype='int32')
assert loaded_im_array.shape == original_grayscale_32bit_array.shape
assert np.all(loaded_im_array == original_grayscale_32bit_array)
# Test down-sampling with bilinear interpolation.
loaded_im = utils.load_img(filename_rgb, target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 3)
loaded_im = utils.load_img(filename_rgba,
color_mode='rgba',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 4)
loaded_im = utils.load_img(filename_rgb,
color_mode='grayscale',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_8bit,
color_mode='grayscale',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_16bit,
color_mode='grayscale',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im, dtype='int16')
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_32bit,
color_mode='grayscale',
target_size=(25, 25))
loaded_im_array = utils.img_to_array(loaded_im, dtype='int32')
assert loaded_im_array.shape == (25, 25, 1)
# Test down-sampling with nearest neighbor interpolation.
loaded_im_nearest = utils.load_img(filename_rgb,
target_size=(25, 25),
interpolation="nearest")
loaded_im_array_nearest = utils.img_to_array(loaded_im_nearest)
assert loaded_im_array_nearest.shape == (25, 25, 3)
assert np.any(loaded_im_array_nearest != loaded_im_array)
loaded_im_nearest = utils.load_img(filename_rgba,
color_mode='rgba',
target_size=(25, 25),
interpolation="nearest")
loaded_im_array_nearest = utils.img_to_array(loaded_im_nearest)
assert loaded_im_array_nearest.shape == (25, 25, 4)
assert np.any(loaded_im_array_nearest != loaded_im_array)
loaded_im = utils.load_img(filename_grayscale_8bit,
color_mode='grayscale',
target_size=(25, 25),
interpolation="nearest")
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_16bit,
color_mode='grayscale',
target_size=(25, 25),
interpolation="nearest")
loaded_im_array = utils.img_to_array(loaded_im, dtype='int16')
assert loaded_im_array.shape == (25, 25, 1)
loaded_im = utils.load_img(filename_grayscale_32bit,
color_mode='grayscale',
target_size=(25, 25),
interpolation="nearest")
loaded_im_array = utils.img_to_array(loaded_im, dtype='int32')
assert loaded_im_array.shape == (25, 25, 1)
# Check that exception is raised if interpolation not supported.
loaded_im = utils.load_img(filename_rgb, interpolation="unsupported")
with pytest.raises(ValueError):
loaded_im = utils.load_img(filename_rgb,
target_size=(25, 25),
interpolation="unsupported")
# Check that the aspect ratio of a square is the same
filename_red_square = str(tmpdir / 'red_square_utils.png')
A = np.zeros((50, 100, 3), dtype=np.uint8) # rectangle image 100x50
A[20:30, 45:55, 0] = 255 # red square 10x10
red_square_array = np.array(A)
red_square = utils.array_to_img(red_square_array, scale=False)
red_square.save(filename_red_square)
loaded_im = utils.load_img(filename_red_square,
target_size=(25, 25),
keep_aspect_ratio=True)
loaded_im_array = utils.img_to_array(loaded_im)
assert loaded_im_array.shape == (25, 25, 3)
red_channel_arr = loaded_im_array[:, :, 0].astype(np.bool)
square_width = np.sum(np.sum(red_channel_arr, axis=0))
square_height = np.sum(np.sum(red_channel_arr, axis=1))
aspect_ratio_result = square_width / square_height
# original square had 1:1 ratio
assert aspect_ratio_result == pytest.approx(1.0)
