Embed Me If You Can: A Geometric Perceptron

This repository is the official implementation of the "Embed Me If You Can: A Geometric Perceptron" ICCV 2021 paper.

[arXiv] [video] [bibtex]

Poster

(best viewed as an image in a new tab)

The proposed MLGP model

Requirements

📋We achieved the original results with Python 3.6.5, torch==1.2.0+cu92, scikit-learn==0.19.1, scipy==1.4.1, numpy==1.15.0, and matplotlib==3.0.3, but we needed to relax the requirements to facilitate the installation.

To install the requirements, run:

pip install -r requirements.txt

Demo

The mlgp_demo.ipynb notebook demonstrates the training and evaluation of our MLGP model and the analysis and visualization of its hidden units.

Training

To train the model(s) in the paper, run the following command:

python train.py 

📋Uncomment specific lines in train.py to use various models described in the paper (default are original hyperparameters). Adjust the get_tetris_data function arguments accordingly.

Evaluation

To evaluate one of the trained models on the corresponding test dataset, run:

python eval.py

📋 Depending on the choice of a trained model, modify the MODEL_PATH variable and the create_test_set function arguments in the eval.py script (examples are provided).

Pre-trained models

You can find the pre-trained models in the pretrained_models folder.

Results

The performances of the models on the test data and in all experiments are presented in Table 1.

📋Use train.py script to train the models with the provided seeds. Use eval.py to evaluate the models on the corresponding test sets.

Citation

@InProceedings{Melnyk_2021_ICCV,
    author    = {Melnyk, Pavlo and Felsberg, Michael and Wadenb\"ack, M\r{a}rten},
    title     = {Embed Me if You Can: A Geometric Perceptron},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {1276-1284}
}