This is the code of the client side of HAPI, a processing system for transfer learning that spans the compute and the object storage tiers. HAPI is presented in the paper "Accelerating Transfer Learning with Cloud Object Stores", co-authored by Arsany Guirguis, Florin Dinu, Do Le Quoc, Javier Picorel, and Rachid Guerraoui. HAPI is built on top of OpenStack swift, the open-source object storage, and PyTorch for ML computations.
HAPI is composed of two components: server and client. In our experiments, we test with 2 machines, each host a separate component.
Please use the server guide to set up the server and the client guide to set up the client.
After a successful setup of the server, run rebuildswift to start it. You can run swift-init all stop to stop it.
To run the HAPI server outside of the proxy server (refer to the paper for more information), run on the server machine:
python3 $HOME/swift/swift/proxy/mllib/server.py
Important: make sure to update the server IP in this line.
On the client side, you can run the following examples:
#split experiment
cd $HOME/collabml_client/application_layer; python3 main.py --dataset imagenet --model myalexnet --num_epochs 1 --batch_size 8000 --freeze --freeze_idx 17 --use_intermediate; cd -
#vanilla experiment
cd $HOME/collabml_client/application_layer; python3 main.py --dataset imagenet --model alexnet --num_epochs 1 --batch_size 8000 --freeze --freeze_idx 17; cd -
#swift-only baseline (called ALL_IN_COS in our paper)
cd $HOME/collabml_client; python3 mlswift_playground.py --dataset imagenet --model alexnet --num_epochs 1 --batch_size 1000 --task training --freeze_idx 17; cd -
The scripts to reproduce the results and the plots of our paper can be found in the experiments directory. Please see run_exp.py for all scripts that we use to run the paper experiments and main.py for the plotting scripts.
CLI documentation can be found here and the client Service API documentation is here.
If you have any question, do not hesitate to contact me on arsany.guirguis@epfl.ch