Skip to content

MODEL_ZOO.md

Latest commit

 

History

History
68 lines (43 loc) · 1.91 KB

MODEL_ZOO.md

File metadata and controls

68 lines (43 loc) · 1.91 KB

Model Zoo

This document lists the available CodeBind checkpoints and explains how to download and place them for evaluation.

Checkpoint Directory

All downloaded checkpoints should be placed under (or connected via symoblic link):

~/.checkpoints/

Available Checkpoints

Model Backbone Download Notes
CodeBind-IB ImageBind Hugging Face CodeBind checkpoint trained from ImageBind features.
ImageBind-Huge ImageBind Meta download Required ImageBind base checkpoint.

Download From Hugging Face

The CodeBind checkpoints are hosted at:

https://huggingface.co/zykev/CodeBind

You can download them with huggingface-cli:

pip install -U huggingface_hub
huggingface-cli download zykev/CodeBind --local-dir ~/.checkpoints/

After downloading, make sure the checkpoint folders are available under ~/.checkpoints/. For example:

ls ~/.checkpoints/asa_lora_MVQ1024x8_256x8

Download ImageBind Base Checkpoint

The ImageBind base checkpoint should also be placed under ~/.checkpoints/:

mkdir -p ~/.checkpoints
wget https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth -O ~/.checkpoints/imagebind_huge.pth

Usage

Evaluation scripts load trained checkpoints through --resume_expname. Point this argument to the downloaded experiment directory under ~/.checkpoints/.

For example:

python main.py \
  --resume_expname ~/.checkpoints/CodeBind/path/to/experiment \

For the provided bash scripts, update the corresponding --resume_expname value in scripts/eval/*.sh to the checkpoint directory you want to evaluate.

Notes