geo-tell/eanet

eanet's Issues

Environment

Hi! I want to know what is your pytorch version.

loss

Hi! I use your project to train my own dataset,but my miou is nan.And loss is almost unchanged after dropping to 2.8 .
Any help will be appreciated ！Thank you

Edge loss didn't converge

Hi @geovsion ,
I just added edge loss (EdgeLoss module) in your code to the original cross entropy loss of a semantic segmentation network (DDCM-Net) , then trained the network on the ISPRS Vaihingen and Potsdam dataset, but the edge loss part didn't converge and kept oscillating like this:

I visualize the edge map in the forward method of EdgeLoss module as follows:
`def forward(self, logits, label):
prediction = F.softmax(logits, dim=1)
edge_maps = []
# kernel size of Laplace filter
ks = 2 * self.radius
filt1 = torch.ones(1, 1, ks, ks)
filt1[:, :, self.radius:2self.radius, self.radius:2self.radius] = -8
filt1.requires_grad = False
filt1 = filt1.cuda()
label = label.unsqueeze(1)
# apply 2D convolution over label using filt1 (Laplacian filter) to get label edge map
lbedge = F.conv2d(label.float(), filt1, bias=None, stride=1, padding=self.radius)
lbedge = 1 - torch.eq(lbedge, 0).float()

    filt2 = torch.ones(self.n_classes, 1, ks, ks)
    filt2[:, :, self.radius:2*self.radius, self.radius:2*self.radius] = -8
    filt2.requires_grad = False
    filt2 = filt2.cuda()
    # get edge map of prediction
    prededge = F.conv2d(prediction.float(), filt2, bias=None,
                        stride=1, padding=self.radius, groups=self.n_classes)
    
    # squash C-dimensional (C is n_classes) edge map to single edge map
    norm = torch.sum(torch.pow(prededge,2), 1).unsqueeze(1)
    prededge = norm/(norm + self.alpha)
    
    
    # mask = lbedge.float()
    # num_positive = torch.sum((mask==1).float()).float()
    # num_negative = torch.sum((mask==0).float()).float()

    # mask[mask == 1] = 1.0 * num_negative / (num_positive + num_negative)
    # mask[mask == 0] = 1.5 * num_positive / (num_positive + num_negative)

    # cost = torch.nn.functional.binary_cross_entropy(
        # prededge.float(),lbedge.float(), weight=mask, reduce=False)
    # return torch.mean(cost)

    
    # visualize edge maps
    edge_maps.append(prededge.float())
    edge_maps.append(lbedge.float())
    edge_maps = torch.cat(edge_maps, dim=0)
    edge_maps_grid = torchvision.utils.make_grid(edge_maps, nrow=prededge.size(0))
    plt.imshow(edge_maps_grid.permute(1, 2, 0).detach().cpu(), cmap='gray')
    plt.title('Edge maps of 1-channel pred and gt')
    plt.show()
    
    return BinaryDiceLoss()(prededge.float(),lbedge.float())`

The visualized result on the Vaihingen dataset is like this:

The optimizer is Adam with AMSGrad. May you help me find the reason? Thanks!