weiaicunzai/pytorch-cifar100

Tensor sizes mismatches

tehreemnaqvi opened this issue · 4 comments

tehreemnaqvi commented

Hi,

I am implementing ResNet18 models using the CIFAR-100 dataset.

I checked all the dimensions but got this error:

RuntimeError: The size of tensor a (100) must match the size of tensor b (32) at non-singleton dimension 3

Can you please tell me how to fix it?

Thanks

weiaicunzai commented

RuntimeError: The size of tensor a (100) must match the size of tensor b (32) at non-singleton dimension 3

Could you plz tell me you are implementing ResNet18 yourself or using my code?Last time I checked, my implementation works fine.

JasperMorrison commented

own dataset, img size 128x128, resnet18 is ok, but vgg16, I got:
RuntimeError: size mismatch, m1: [128 x 8192], m2: [512 x 4096]

weiaicunzai commented

[128 x 8192], m2: [512 x 4096]

VGG does not have adaptive pooling layer, so you have to modify the fully connected layer in VGG16 to adapt your dataset since cifar100 only have 32x32 image size.

  • 👍1
jaeho3690 commented

You have to check the dimension output in the vgg.py line 38. After checking the output dimension, you should modify the self.classifier function that matches your size.