vislearnlabpy 0.0.4.3

Visual Learning Lab utility files and pipelines

vllpy

This is a package with common utility functions, files and pipelines for the Visual Learning Lab. This package uses python>=3.11.

It is recommended you create a conda environment to start using this package but this step is optional. To do so, run the commands below:

conda create -n vislearnlabpy python=3.12
conda activate vislearnlabpy

Then, activate the environment and simply install vislearnlabpy and CLIP by running the commands below in your terminal.

pip install --upgrade vislearnlabpy
pip install git+https://github.com/openai/CLIP.git

You may also have to install PyTorch manually, ensuring that you have the right version but the right version may also be installed by default with CLIP. To install the right version of PyTorch on the Tversky server, run:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Here is as an example of how to generate a CSV file with embeddings from a list of images in a directory. You can also use this to generate npy files and doc files by changing the output type in the command below, and generate the embeddings from a CSV file instead by using input_file instead of input_dir

python embedding_generator.py --input_dir examples/input --output_path examples/output --output_type csv --overwrite --model_name clip

Here is an example of how to generate npy files for embeddings from a list of images within subdirectories in a folder on Tversky using the dinov3-babyview model (other available models include dinov3, dinov2, dino_say_vitb14, dino_imagenet100_vitb14 and clip). For certain models, you will need a HuggingFace account and a HuggingFace access token. Set HF_TOKEN=xxxx in a .env file in the main directory of your project to be able to access restricted models. Also note that by default, npy files are generated in parallel using Ray with all available GPUs. To use specific GPUs, like the first and second GPUs for example, try using export CUDA_VISIBLE_DEVICES=0,1 Ideally run this command in a tmux window if you have a large input directory.

python embedding_generator.py --input_dir /labs/vislearnlab/data/THINGS-dataset/object_images --output_path /labs/vislearnlab/data/THINGS-dataset/things_embeddings --output_type npy --overwrite --model_name dinov3-babyview --subdirs --batch_size 1000 --ray_temp_dir /Scratch/tmp/ray

For more detailed examples, please look at the demo in the Jupyter notebooks within the examples folder.