Skip to content

Latest commit

 

History

History
15 lines (9 loc) · 1.3 KB

README.md

File metadata and controls

15 lines (9 loc) · 1.3 KB

This is the codebase to reproduce the results in the paper "Natural Image Statistics and Modeling Neural Representations"

Reproducing the results:

You have the option of either reproducing each step yourself or there are download scripts provided at each step for downloading our results. There are n steps to reproducing. At each step, cd into the corresponding directories - train_simclr, generate_activations and rep_analysis:

  1. Install the conda env provided using conda env create -f neural_pred.yml. Then activate the environment using conda activate domain_rep

  2. Download the neural datasets needed for the pipeline using python download_neural_datasets.py -d [DATASETS] -o [OUTPUT_DIR] See python download_neural_datasets -h for more info.

  3. Train the simclr model on whatever datasets you want using python train.py [DATASET_PATH] --options. Currently all the flags are set to the hyperparameters we used but if you'd like to change them, see python train_simclr/train.py -h.

  4. Generate model activations to each dataset using python generate_model_activations.py -d [dataset] -m [model names] -c [checkpoint paths]. See python generate_model_activations.py

  5. Once you've generated the model activations, you can use any of the scripts located in rep_analysis to reproduce our analysis.