size mismatch for aspp when loading pre-trained model

[LindaSt]

Hi!

Very nice repo! I'm currently trying to integrate your model into our framework (https://github.com/DIVA-DIA/DeepDIVA, feel free to check it out!). However, when I load the provided weights for deeplabv3 I get the following error:

    size mismatch for aspp.conv_1x1_4.weight: copying a param with shape torch.Size([20, 256, 1, 1]) from checkpoint, the shape in current model is torch.Size([8, 256, 1, 1]).
    size mismatch for aspp.conv_1x1_4.bias: copying a param with shape torch.Size([20]) from checkpoint, the shape in current model is torch.Size([8]). (deeplabv3.py:50)

I am using exactly the Resnet (ResNet18_OS8) and the ASPP (no bottleneck) that you are using in your code. Do you know what could be causing this?

Thank you very much already in advance.

Cheers, Linda

[fregu856]

Hi,

20 is the number of classes i trained the model for (self.num_classes = 20 in https://github.com/fregu856/deeplabv3/blob/master/model/deeplabv3.py), it seems like you have set this to 8 instead?

Regards

// Fredrik

[LindaSt]

I think I just realized what the problem is. I am trying to do transfer learning and have specified a different number of classes than in your dataset. That is what is causing the mismatch.

[fregu856]

Yes, exactly! You could probably load the pre-trained model into a placeholder model with 20 classes, then loop through the parameters (but skip aspp.conv_1x1_4.weight and aspp.conv_1x1_4.bias) and copy their values to your model.

[LindaSt]

Thank you so much for your quick response! Yes, my model has a different number of classes.

Somewhat unrelated question: the pre-trained weights provided for the resnets are for the cityscape model aswell, or from something else?

[fregu856]

The pre-trained resnets are taken straight from torchvision, i.e., downloaded from the urls at the top of https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py.