When pytorch was upgraded from 0.4.1 to 1.1.0 and torchvision was upgraded from 0.2.0 to 0.3.0, the training time was changed from the original 60min to 90min. (Ubuntu16.04, cuda9.0, gcc5.4.0, 1080Ti)
Which model and dataset are you using?
I use the seq2seq model which contains ResNet and RNN and the dataset is Synthetic text image dataset (~more than 10 million images).
Could you update to the latest stable PyTorch version (
1.5) and rerun the code please?