def transform_train(self, img, *targets):
for idim in range(len(self.dims)):
if _np.random.random() < 0.5:
img = _u.flipany(img, self.dims[idim])
if self.tgtdims is not None:
targets = [
_u.flipany(t, self.tgtdims[idim]) for t in targets
]
return img, targets
def transform_pred(self, img, i, fast):
assert i < self.npreds(fast), "This should never happen, please file an issue."
for idim, d in enumerate(self.dims):
if i >> idim & 1:
img = _u.flipany(img, d)
return img
def transform_pred(self, img, i, fast):
assert i < self.npreds(
fast), "This should never happen, please file an issue."
for idim, d in enumerate(self.dims):
if i >> idim & 1:
img = _u.flipany(img, d)
return img
def flipped_classes(X, y, n, le, old, new):
"""
Horizontally flips all images in `X` which are labeled as `old` and label them as `new`.
Returns the flipped X, y, n.
"""
indices = np.where(y == le.transform(old))[0]
return (flipany(X[indices], dim=3),
np.full(len(indices), le.transform(new),
dtype=y.dtype), tuple(n[i] for i in indices))
def flipped_classes(X, y, n, le, old, new):
"""
Horizontally flips all images in `X` which are labeled as `old` and label them as `new`.
Returns the flipped X, y, n.
"""
indices = np.where(y == le.transform(old))[0]
return (
flipany(X[indices], dim=3),
np.full(len(indices), le.transform(new), dtype=y.dtype),
tuple(n[i] for i in indices)
)
def flipall_images(images):
"""
Horizontally flips all given `images`, assuming `images` to be a list of HWC tensors.
"""
return [flipany(img, dim=1) for img in images]
def flipped(X, y, n, old, new):
indices = np.where(y == old)[0]
return flipany(X[indices],
dim=3), np.full(len(indices), new,
dtype=y.dtype), [n[i] for i in indices]
def transform_train(self, img, *targets):
for d in self.dims:
if _np.random.random() < 0.5:
img = _u.flipany(img, d)
return img, targets
def flipall_images(images):
"""
Horizontally flips all given `images`, assuming `images` to be a list of HWC tensors.
"""
return [flipany(img, dim=1) for img in images]
def flipped(X, y, n, old, new):
indices = np.where(y == old)[0]
return flipany(X[indices], dim=3), np.full(len(indices), new, dtype=y.dtype), [n[i] for i in indices]
