is there a way to run this project without using CUDA?
rohith46 opened this issue · 6 comments
I have installed docker image of decaNLP on my local and want to try custom dataset(Json dict file: context, questions, answers). I don't have Nvidia hardware to install CUDA. so using docker instead of Nvidia-docker (( docker run -it --rm -v pwd:/decaNLP/ decanlp bash -c "python /decaNLP/predict.py --evaluate valid --path /decaNLP/mqan_decanlp_qa_first --checkpoint_name model.pth --tasks my_custom_dataset" ))
Is there a possibility to try model.pth (pre-trained model) to predict on the custom dataset without CUDA dependency?
Yes there should be; I'll upload a separate Dockerfile/image for CPU usage, and I might need to reupload the model checkpoints if they were saved on GPU.
Hi,
I'm trying to do the same. Is it possible to run the default model without using the CUDA? Is the CPU version uploaded? Thanks
Should be possible on https://github.com/salesforce/decaNLP/tree/pytorch_upgrade_041, which I'll merge ASAP. The migration somehow increased my GPU memory usage, so I'm holding off on merging until I figure out why that is. In the meantime, you can use that branch to run on CPU with new prebuilt Docker images. You will probably need to use a smaller --train_batch_tokens though.
I experimented with the old pretrained GPU models on a compute cluster and they worked great! However, I'm interested now in running a server instance of decaNLP on a machine with no GPUs. The new branch looks like it covers CPU support well, but I was unable to download the new CPU pretrained models (didn't have a problem with the previous models).
It looks like I was able to successfully connect to your S3 bucket, but I got an ERROR 403: Forbidden. If it's due to a similar circumstance as this, then it should just be a matter of fixing permissions on the uploads.
Thanks for all your work on this -- this is one of the best research paper code bases I've worked with.
Thanks for pointing that out! Don’t have access to the permissions right now, but I’ll update when I’ve confirmed they are okay.
Should be updated now. Let me know (and reopen) if not!