Commit 903c4dc7 authored by Daniil Pastukhov's avatar Daniil Pastukhov 🏳🌈


parent 4f388d5e
# Petfinder (kaggle competition) solution
- Link to the competition -
- Link to the competition - (data can be downloaded there)
- Semestral work for NI-MVI course (FIT CTU)
- Runs at -
## Tech stack
- Python 3.7, PyTorch, PyTorch Lightning, Albumentations ([github](
- Python 3.7, PyTorch 1.10.0, PyTorch Lightning, Albumentations ([github](
## Approaches
1. Use just images and CNN as a regressor to predict *pawpularity* score.
2. Same as (1), but the problem can be interpreted as classification problem since all targets are within 1-100 range.
3. Use auxiliary metadata to improve predictions. LightGBM/CatBoost can be utilized.
4. Use pre-trained CNN (ResNet, EfficientNet, etc.) as a backbone.
5. Employ img2vec model to extract features from the images, then traning a model using them.
6. Use SOTA models (e.g., Swin transformer ([github](, [arxiv](
7. Ensembling via stacking.
## Run
- Repo contains 2 Jupyter notebooks that contain CNN and EfficientNet solutions.
## Problems
- SOTA models are very hard to train (even if they are pre-trained). My current setup (GTX 1650S) cannot handle such a big NN.
- Even 3-4 conv layers in CNN are slow to train :-(
## Working on
- Try CNN as a regressor.
- Try CNN as a classificator.
- Try pre-trained nets (EfficientNet, etc).
- Img2vec -> tree boosting.
## Notes
- A competition has a limited resources that can be used for the submission so solutions in the repo are slightly modifed (they utilzie metadata), but they couldn't be submitted to the competition.
\ No newline at end of file
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment