/NeuralDataServer-FastAdapt

Primary LanguagePython

NeuralDataServer-FastAdapt

This repo contains standalone scripts for the fast adaptation module that can be used off-the-shelf to generate transfer performance.

Prerequisite

Running the scripts assumes a directory [PATH/TO/CLIENT/DATASET] containing images of the client dataset, and a directory [PATH/TO/DOWNLOADED/EXPERT/MODELS] containing pre-trained weights of expert models.

conda create --name fast-adapt
conda install pytorch torchvision
conda install torchnet

Usage

cd src
python main.py --imagedir [PATH/TO/CLIENT/DATASET] --experts_dir [PATH/TO/DOWNLOADED/EXPERT/MODELS]

Running it off-the-shelf will generate a z.pickle file inside experiments/GenericFastAdapt/ containing the performance metric for each expert. Send this file back to the server to obtain a subset of server data that is most relevant for images inside [PATH/TO/CLIENT/DATASET].