Skip to content

Latest commit

 

History

History
113 lines (74 loc) · 3.37 KB

REPRODUCIBILITY.md

File metadata and controls

113 lines (74 loc) · 3.37 KB
Raw

Reproducing the results

Prerequisites

  • Python version 3.10 (preferable).
  • Download datasets as instructed below.
  • Download pretrained checkpoints as instructed below.

You can download a copy of all the files in this repository by cloning the git repository:

git clone https://github.com/naamiinepal/synthetic-boost

or downloading a zip archive from GitHub for the latest commit.

All source code for the paper are in the src folder and the related scripts are in scripts folder. The Jupyter notebooks used for the evalutation and visualization are in notebooks folder.

The LaTeX sources for the manuscript text and figures are in paper folder.

Setting up your environment

Setting up Python

You'll need a working Python 3 environment to run the code. Although we used version 3.10 to produce the results in the paper, the code should work with any Python 3.8+ version.

Setting up Virtual Environment

Assuming you have setup required python version, you can create a virtual environment using python's built-in venv module.

    python -m venv .venv

python is the path to your python executable. You can find it by running which python or which python3 in your terminal.

This will create a .venv folder in your current directory. You can activate the virtual environment by running:

    source .venv/bin/activate

Installing dependencies

Once you have activated the virtual environment, you can install the required dependencies by running:

    pip install -r requirements.txt

Downloading the data

You need to download the data from the following sources:

Downloading Models

CLIPSeg

CLIPSeg are automatically downloaded when you run the scripts from Huggingface Model Hub.

CRIS

Downloading CLIP

The Resnet 50 CLIP model needed for CRIS can be downloaded from here using the following command.

    wget https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt -O pretrain/RN50.pt

Downloading CRIS

We have not yet found the official link to download the CRIS model. We have used the model from CRIS's repo's issue using this OneDrive link.

Converting CRIS Model

The downloaded CRIS model needs to be loaded in DataParallel mode. To convert the model, run the following command after saving the downloaded model to pretrain/cris_best.pt:

    python scripts/convert_cris_model.py

Training VLSMs

CAMUS

Finetune the VLSMs with CAMUS dataset:

    bash scripts/camus.sh

SDM CAMUS

Finetune the VLSMs with SDM CAMUS dataset:

    bash scripts/sdm_camus.sh

PT-FT Strategy

Finetune the VLSMs with CAMUS dataset, already finetuned on SDM CAMUS dataset:

    bash scripts/pt_ft.sh