env.md

Setting up a development environment

If you do not have conda installed, you can install it from here. This code has been tested on Linux.

Create a new conda environment with Python 3.8.8:

conda create -n testoftime python=3.8.8 -y

Install the required packages:

0. Activate the environment:

conda activate testoftime

1. Install PyTorch 1.12.1 with CUDA 11.1 support:

pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116

Depending on your CUDA version, you may have to install this differently. See here for more details.

Note: Before proceeding, make sure to add $PWD to the PYTHONPATH environment variable:

export PYTHONPATH=$PWD:$PYTHONPATH

2. Setup code for VideoCLIP model: We borrow code for VideoCLIP model from the official repository. In this repo, we provide an adaptation of their code in external/fairseq/examples/MMPT. The credit for the original code solely lies with the authors of the original repository at FAIR.

   2.1: Download best VideoCLIP checkpoint:

   ckpt_dir=external/requirements/fairseq/runs/retri/videoclip
   mkdir -p $ckpt_dir
   wget https://dl.fbaipublicfiles.com/MMPT/retri/videoclip/checkpoint_best.pt -O $ckpt_dir/checkpoint_best.pt
   

   2.2: Check that loading the model works:

   python external/utils_videoclip.py
   

3. Install the rest of the required packages (in this order):

pip install ipdb
pip install transformers==3.4
pip install protobuf==3.20.0
pip install omegaconf==2.0.6
pip install termcolor==2.2.0
pip install pytorch_lightning==1.7.6
pip install wandb==0.13.3
pip install editdistance==0.6.2
pip install pandas==1.5.3

3. Check that the installation was successful:

python setup/check_packages.py