def __mul__(self, other):
if isinstance(other, Attribute):
self_data, other_data = self.data, other.data
# Repeat along missing colour dimensions if necessary
if self.data.ndim < other.data.ndim:
self_data = self.data.unsqueeze(0).repeat(
other.data.shape[0], 1, 1)
elif self.data.ndim > other.data.ndim:
other_data = other.data.unsqueeze(0).repeat(
self.data.shape[0], 1, 1)
# Compare shape of data tensors
if self_data.shape != other_data.shape:
self_data = denorm(self_data)
other_data = denorm(other_data)
if self_data.size > other_data.size:
other_data = other_data.resize(self_data.size, resample=2)
else:
self_data = self_data.resize(other_data.size, resample=2)
self_data = norm(self_data, unsqueeze=False, grad=False)
other_data = norm(other_data, unsqueeze=False, grad=False)
return Attribute(self_data * other_data, self.input_data)
elif isinstance(other, (int, float)):
return Attribute(self.data * other, self.input_data)
else:
raise ValueError(f"Can't multiply by type {type(other)}")
def test_norm_denorm():
img = Image.fromarray(np.random.random((32, 32, 3)), 'RGB')
data = norm(img)
denorm_data = denorm(data, image=False)
assert np.isclose(np.array(img), denorm_data).all()
def test_denorm():
data = torch.randn(1, 3, 50, 50)
denorm_data = denorm(data, image=False)
assert denorm_data.max() <= 255
assert denorm_data.min() >= 0
assert denorm_data.shape == (50, 50, 3)
def show(self,
ax=None,
show_image=True,
alpha=0.4,
cmap='magma',
colorbar=False,
**kwargs):
"""Show the generated attribution map.
Parameters:
show_image (bool): show the denormalised input image overlaid on the heatmap.
ax: axes on which to plot image.
colorbar (bool): show a colorbar.
cmap: matplotlib colourmap.
alpha (float): transparency value alpha for heatmap.
kwargs: passed to `denorm`. Used to change mean and std normalization values.
"""
if ax is None:
_, ax = plt.subplots()
sz = list(self.input_data.shape[2:])
if show_image:
input_image = denorm(self.input_data[0], **kwargs)
ax.imshow(input_image)
data = self.data
if (data < 0).any():
data = (data - data.min()) / (data.max() - data.min())
if data.ndim >= 3:
data = data.squeeze()
if data.ndim == 3:
data = data.permute(1, 2, 0)
else:
raise RuntimeError(
f"Can't display data shape {self.data.shape} as an image.")
im = ax.imshow(data,
alpha=alpha,
extent=(0, *sz[::-1], 0),
interpolation='bilinear',
cmap=cmap)
if colorbar:
ax.figure.colorbar(im, ax=ax)
def _repr_html_(self):
from IPython.display import display
return display(denorm(self.image))