I have trained and save model with python 3.7 and pytorch 1.0.1.post2. When I load it on the same machine, it works fine.
However, when I tried to load it on the other machine which has python 3.6 and pytorch 1.0.1 installed, the results were completely off (loss was significantly larger). After updating python and pytorch, everything works as expected.
Is such behavior expected?
Both machines have the same cuda version but different GPUs (1080 Ti and 2080 Ti). Code is absolutely identical as well as checkpoint file.