def forward(self, pred, target):
assert pred.dim() == target.dim(), "inconsistent dimensions"
mask = (target>0).detach()
#mask = (target > 0) & (target <= 500).detach()
#pred = normalize_prediction_robust(pred,mask)
#target = normalize_prediction_robust(target,mask)
res = pred - target
res = res[mask.bool()].abs().mean()
M = torch.sum(mask)
self.loss = res
gradl = 0
#pred = normalize_prediction_robust(pred,mask)
#target = normalize_prediction_robust(target,mask)
for scale in range(1, 5):
pred_r = F.interpolate(pred, scale_factor=1/scale, mode='bilinear')
target_r = F.interpolate(target,scale_factor=1/scale, mode='bilinear')
mask = (target_r > 0).detach()
pgrads: torch.Tensor = kornia.spatial_gradient(pred_r, order=1)
tgrads: torch.Tensor = kornia.spatial_gradient(target_r, order=1)
#res = pred_r - target_r
#trim = 0.2
diff2 = (pgrads[:, :, 0] - tgrads[:, :, 0]) + (pgrads[:, :, 1] - tgrads[:, :, 1])
#res = res[mask.bool()].abs().mean()
#diff2 = diff2[mask.bool()].abs().mean()
gradl+= diff2.abs().mean()
# trimmed, _ = torch.sort(res.view(-1), descending=False)[
# : int(len(res) * (1.0 - trim))
# ]
#self.loss = trimmed.sum() / (2 * M.sum()) + diff2
self.loss+= 0.5 * gradl
return self.loss
def imshow(input: torch.Tensor):
out: torch.Tensor = torchvision.utils.make_grid(input, nrow=2, padding=1)
out_np: np.array = kornia.tensor_to_image(out)
plt.imshow(out_np)
plt.axis('off')
#############################
# Show original
imshow(x_rgb)
#################################
# Compute the 1st order derivates
grads: torch.Tensor = kornia.spatial_gradient(x_gray, order=1) # BxCx2xHxW
grads_x = grads[:, :, 0]
grads_y = grads[:, :, 1]
#################################
# Show first derivatives in x
imshow(grads_x)
#################################
# Show first derivatives in y
imshow(grads_y)
#################################
# Sobel Edges
# Once with the gradients in the two directions we can computet the Sobel edges.
# However, in kornia we already have it implemented.