i want to perform without transform pipeline.
IS there a function taht can be directly called to do zoom overn image of any size.
Here is what i have made by changing some git hub code. Provide your views on this
It behaves like Fastai one. but the problem is i get Meomory allocate issue.
@muellerzr any opinion of yours on this.
def transform_matrix_offset_center(matrix, x, y):
"""Apply offset to a transform matrix so that the image is
transformed about the center of the image.
NOTE: This is a fairly simple operaion, so can easily be
moved to full torch.
Arguments
---------
matrix : 3x3 matrix/array
x : integer
height dimension of image to be transformed
y : integer
width dimension of image to be transformed
"""
o_x = float(x) / 2 + 0.5
o_y = float(y) / 2 + 0.5
offset_matrix = np.array([[1, 0, o_x], [0, 1, o_y], [0, 0, 1]])
reset_matrix = np.array([[1, 0, -o_x], [0, 1, -o_y], [0, 0, 1]])
transform_matrix = np.dot(np.dot(offset_matrix, matrix), reset_matrix)
return transform_matrix
def apply_transform(x, transform, fill_mode='nearest', fill_value=0.):
"""Applies an affine transform to a 2D array, or to each channel of a 3D array.
NOTE: this can and certainly should be moved to full torch operations.
Arguments
---------
x : np.ndarray
array to transform. NOTE: array should be ordered CHW
transform : 3x3 affine transform matrix
matrix to apply
"""
x = x.astype('float32')
transform = transform_matrix_offset_center(transform, x.shape[1], x.shape[2])
final_affine_matrix = transform[:2, :2]
final_offset = transform[:2, 2]
channel_images = [ndi.interpolation.affine_transform(x_channel, final_affine_matrix,
final_offset, order=0, mode=fill_mode, cval=fill_value) for x_channel in x]
x = np.stack(channel_images, axis=0)
return x
class Zoom(object):
def __init__(self,
zoom_range,
fill_mode='constant',
fill_value=0,
target_fill_mode='nearest',
target_fill_value=0.,
lazy=False):
"""Randomly zoom in and/or out on an image
Arguments
---------
zoom_range : tuple or list with 2 values, both between (0, infinity)
lower and upper bounds on percent zoom.
Anything less than 1.0 will zoom in on the image,
anything greater than 1.0 will zoom out on the image.
e.g. (0.7, 1.0) will only zoom in,
(1.0, 1.4) will only zoom out,
(0.7, 1.4) will randomly zoom in or out
fill_mode : string in {'constant', 'nearest'}
how to fill the empty space caused by the transform
fill_value : float
the value to fill the empty space with if fill_mode='constant'
lazy : boolean
if true, perform the transform on the tensor and return the tensor
if false, only create the affine transform matrix and return that
"""
if not isinstance(zoom_range, list) and not isinstance(zoom_range, tuple):
raise ValueError('zoom_range must be tuple or list with 2 values')
self.zoom_range = zoom_range
self.fill_mode = fill_mode
self.fill_value = fill_value
self.target_fill_mode = target_fill_mode
self.target_fill_value = target_fill_value
self.lazy = lazy
def __call__(self, x, y=None):
zx = np.random.uniform(self.zoom_range[0], self.zoom_range[1])
zy = np.random.uniform(self.zoom_range[0], self.zoom_range[1])
zoom_matrix = np.array([[1/zx, 0, 0],
[0, 1/zy, 0],
[0, 0, 1]])
if self.lazy:
return zoom_matrix
else:
x_transformed = torch.from_numpy(apply_transform(x.numpy(),
zoom_matrix, fill_mode=self.fill_mode, fill_value=self.fill_value))
if y:
y_transformed = torch.from_numpy(apply_transform(y.numpy(), zoom_matrix,
fill_mode=self.target_fill_mode, fill_value=self.target_fill_value))
return x_transformed, y_transformed
else:
print(x_transformed.size())
return x_transformed
zoom_fn=Zoom((1,1.2))