The competition has an objective of image classification in experimantal noise of biological signals. Here the proposed algorithms detects different genetic perturbations.
- GPU: 1xTesla K80
- PyTorch, albumentations
As the cellular images have origin from 4 types of experiments (HEPG2, HUVEC, RPE, U2OS) we have trained 4 different models in parallel for each experiment and then concatenated the predictions.
The solution represents:
- models:
- EfficientNet-B0
- augmentations:
- Albumentations library
- Rotate90, HorizontalFlip, Brightness, Contrast, ColorJitter
- optimizer: Adam
- loss: CrossEntropyLoss
- batch size: 16