I tried tracing this back and it seems like the variable lr is supposed to be a float, but has been converted in the middle of this function call learn.fit. Is is a problem with the libraries or am I missing some files?
My Mac is running on Mac OS X Sierra 10.12.6, CUDA 8, cuDNN 6, and was implemented with a GTX 1070 locally. I used pip install fastai and did not use Anaconda since I had setup my pyTorch successfully already.
Same here on Ubuntu 17.10, Python 3.6.4, pytorch 0.4.0a0+bc1b4c8.
Python installed via anaconda, pytorch from source, to compile against my older gpu. However, for memory reasons I had to run the code on the cpu
I’m on a Windows laptop, but I’m also having this problem. I made some changes to deal with out-of-memory issues, so I thought it had something to do with that, but apparently others that haven’t made any changes are also having issues. I suppose I should also mention that I just downloaded everything this weekend.
It looks like there were some very recent changes to optim/sgd.py in the pytorch library that might be causing this.
Yep, I think there were some recent changes to the optim/sgd.py library that maybe the current fastai library did not update to. It is possible @jeremy can help us look into this issue (thank you in advance)? I think it just happened recently~
Edit: It seems like people having this problem were all on pytorch 0.4.0a0
I thought I’d add that the reason I am using the newer pytorch is that I had to build from source, as the standard package version won’t work on old-ish GPUs.
I was able to get this to work by just going into optim/sgd.py and removing the if-statement that was choking it up.
