Example training method for Depth Estimation VNL network

[atrah22]

Hello,

The code shows to train for the Semantic Segmentation Task. Is any demo or code available for the Depth Estimation task?

BRs, Atul

[irfanICMLL]

I use this code base, and add the distillation loss on it.

https://github.com/YvanYin/VNL_Monocular_Depth_Prediction

[Rui-Zhou-2]

Hello, when I apply the pixel distillation for depth estimation task, `class CriterionPixelWise(nn.Module): def init(self, ignore_index=255, use_weight=True, reduction=True): super(CriterionPixelWise, self).init() self.ignore_index = ignore_index self.criterion = torch.nn.CrossEntropyLoss(ignore_index=ignore_index, reduction=reduction)

def forward(self, preds_S, preds_T):
    assert preds_S.shape == preds_T.shape,'the output dim of teacher and student differ'
    N,C,W,H = preds_S.shape
    softmax_pred_T = F.softmax(preds_T.permute(0,2,3,1).contiguous().view(-1,C), dim=1)
    logsoftmax = nn.LogSoftmax(dim=1)
    loss = (torch.sum( - softmax_pred_T * logsoftmax(preds_S.permute(0,2,3,1).contiguous().view(-1,C))))/W/H
    return loss / N`

it will all become zero after the softmax

I am so confusing about how to implement this loss on the depth estimation task,

could you give some advice?

Best regards

I use this code base, and add the distillation loss on it.

https://github.com/YvanYin/VNL_Monocular_Depth_Prediction