search

zhirongw / lemniscate.pytorch

Unsupervised Feature Learning via Non-parametric Instance Discrimination
746 stars 134 forks source link

RuntimeError: set_sizes_contiguous is not allowed on Tensor created from .data or .detach(), in Pytorch 1.1.0 #9

Open HHHedo opened 5 years ago

HHHedo commented 5 years ago

My pytorch is 1.1.0. x.data.resize(batchSize, inputSize, 1)_ in NCEAverage.py doesn't work. _with torch.nograd(): x.resize(batchSize, inputSize, 1) still doesn't work. I replaced it as x.unsqueeze(2))_ . I wonder if this function works the same as x.data.resize(batchSize, inputSize, 1)_

ChongjianGE commented 4 years ago

Hi, there. Have you solved the compatibility of Pytorch?

robert780612 commented 4 years ago

Use torch.reshape out = torch.bmm(weight, torch.reshape(x, (batchSize, inputSize, 1)))

Crestina2001 commented 1 year ago

It seems that robert's suggestions can work, but gradOutput.data.resize(batchSize, 1, K+1) in around line 60 should be changed to gradOutput.resize(batchSize, 1, K + 1).