• ***News!!! We introduced a new generic framework to measure similarity between tasks. Please find our code and paper(accepted in ECCV 2020).

CVPR_RSA_Taskonomy

code for CVPR paper "Representation Similarity Analysis for Efficient Task Taxonomy and Transfer Learning"

This repo is build upon the code from the Taskonomy paper.

Step1 RDM creation

1. Switch to taskonomy/taskbank
2. Follow installation instructions and downloading pretrained models
3. Download taskonomy images used for RDM creation here
4. Change current directory to taskonomy/taskbank/tools directory
5. Run python generate_features_taskonomy_500 --img_dir "path to downloaded image directory" --save_dir "path to save features"
6. Run python create_taskonomy_RDMs_enc_final.py --feat_dir "path to saved features" --save_dir "path to save RDMs"

P.S.: Make sure you have enough space as saved features require lot of space (67 GB).

Step2 Download my forked copy of RSA toolbox

Step3 RSA analysis (generate figures from the paper)

1. Change current directory to RSA_analysis
2. run demo.m after modifying the rdm_dir, rsa_dir, and save_dir paths to corresponding paths in the demo.m to generate results for Figure 4 and 5 from the paper

The code to generate results for small models and pascal VOC will be released soon. For early access or any other queries please contact kshitijdwivedi93@gmail.com