I was trying some downloadable datasets from torchvision and found a few differences that are affecting running the pre-trained VGG’s at times, which is due to the way the data tuples are shaped. For example,
Cifra10 has the following shape that works fine with VGG11:
cifar10_train_set.train_data.shape
(50000, 32, 32, 3)
STL10, on the otherhand, has the follwoing shape that does not work with VGG11:
stl10_train_set.data.shape
(5000, 3, 96, 96)
Shouldn’t the two datasets have the same shape?
Do I have to add something like transforms.Lambda(lambda x: x.permute(...)) to correct STL10, as Cifar10 is running well?
The above caused the following error for STL10:
RuntimeError: size mismatch, m1: [20 x 512], m2: [4608 x 4096] at /opt/conda/conda-bld/pytorch_1524584710464/work/aten/src/THC/generic/THCTensorMathBlas.cu:249