Offical Repo of paper Pixel Sentence Representation Learning

• Model Checkpoint: at HuggingFace
• Github Repo: Github Repo
• Paper: https://arxiv.org/pdf/2402.08183.pdf

conda create -n pixel python=3.9 -y && conda activate pixel
git clone https://github.com/gowitheflow-1998/Pixel-Linguist.git

conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
conda install -c conda-forge pycairo pygobject manimpango scikit-learn
cd Pixel-Linguist
pip install -r requirements.txt
pip install -e .

(Do not need to download this if directly git clone our repo)

python scripts/data/download_fallback_fonts.py ‘data/fallback_fonts’

sts benchmark:

python tools/evaluation_sts.py

Adjust the specific language you want to evaluate.

beir:

python tools/evaluation_retrieval.py

It is convenient to evaluate on other datasets available on BEIR, other than Natural Questions that we evaluated in the paper. Simply modify dataset name in the script.

Step 0: Visual alignment step:

bash run_bash/0-run_unsup.sh

For this step, run separately on all our unsup datasets to create 4 checkpoints, and do an ensemble using tools/ensemble.py.

Step 1: Topical alignment step:

bash run_bash/1-run_wikispan.sh

Step 2: Reasoning alignment step:

bash run_bash/2-run_allnli_finetune.sh

Step 3: multilingual transfer step:

bash run_bash/3-run_allnli-pm.sh

Go back and forth between Step 2 and Step 3 (see paper for exact procedure of the iterative training where "leapfrogging" pattern is found!) for 2-3 times for maximum performance, please end the training with English allnli, instead of parallel data.

We find that training with an extra MLP (using PIXELForSequenceClassification class) but do inference without (using PIXELForRepresentation, which then drops the MLP), boosts the semantics performance a little bit, providing the performance in the latest version of the paper.