emi
October 28, 2018, 8:27am
1
I am using a Linux machine here where I manually installed the packages by looking at the bash commands from http://course-v3.fast.ai/start_aws.html#step-6-access-fastai-materials (before that I installed miniconda from https://conda.io/miniconda.html )
I am getting ValueError: padding_mode needs to be 'zeros' or 'border', but got reflection and had to manually patch fastai/vision/image.py and replace reflection with zeros.
Once that is done, I am able to train locally, albeit very slowly as I’m also using pytorch-cpu.
The above error was also independently discovered by @agoldina (https://forums.fast.ai/t/chicago-usa-area/27424/19?u=emi ).
This seems to me to be a packaging problem? Are the conda packages stale?
I’m also confused as to why we are using pytorch-nightly when the baseline is supposed to be pytorch 1.0.
emi
October 28, 2018, 11:08am
2
Nevermind, it seems to be a problem with the packages I installed.
pytorch-cpu is not the CPU backend for pytorch-nightly, it’s an older version of Pytorch actually.
I had to install pytorch-nightly-cpu
1 Like
emi
October 29, 2018, 10:25am
3
For anybody following fastai-1.0.15 also removed ConvLearner you have to downgrade to 1.0.14 with conda install -c fastai fastai=1.0.14
jeremy
October 29, 2018, 1:26pm
4
No don’t do that - instead follow the instructions in the official update thread:
Note that after you conda update and git pull, you’ll find that there’s been one change to fastai - ConvLearner is now called create_cnn. You’ll need to make that change in any notebooks that you’ve created too.
emi
October 29, 2018, 2:13pm
5
Usually 1.0 means a stable release. You should not introduce breaking changes in a dot release.
jianxun
January 28, 2019, 11:43pm
6
I ran into the exact same problem on AWS EC2 even though I had followed your instructions https://course.fast.ai/start_aws.html
