https://github.com/filnow/skin-cancer-detection

repo for training and experiments with skin-cancer dataset from kaggle
https://github.com/filnow/skin-cancer-detection

classification kaggle lightning pytorch torchmetrics wandb

Last synced: 6 months ago
JSON representation

repo for training and experiments with skin-cancer dataset from kaggle

• Host: GitHub
• URL: https://github.com/filnow/skin-cancer-detection
• Owner: filnow
• Created: 2024-05-07T08:48:23.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2024-06-05T18:37:19.000Z (over 1 year ago)
• Last Synced: 2025-05-08T22:58:51.903Z (6 months ago)
• Topics: classification, kaggle, lightning, pytorch, torchmetrics, wandb
• Language: Jupyter Notebook
• Homepage:
• Size: 42 MB
• Stars: 1
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Skin-Cancer-Detection

Repo for training and experiments with skin-cancer dataset from kaggle

[Link to dataset](https://www.kaggle.com/datasets/fanconic/skin-cancer-malignant-vs-benign)

[Link to kaggle notebook](https://www.kaggle.com/code/filnow/acc-92-on-test-set-with-simple-pytorch)

## Experiments

I was able to experiments only with CNN that fits in 4GB of RAM on my GPU.

Diffrent architectures that was use: 

    * Custom CNN - my CNN with 3 conv layers and 2 fc layers

    * MobileNetV2 - pretrained wieghts from ImageNetV2, 3.5MLN PARM

    * RegNet X 400MF - pretrained weights from ImageNetV2, 5.5MLN PARM

    * EfficientNetB0 - pretrained weights from ImageNetV1, 5.3MLN PARM

The best was EfficientNetB0, with early stopping it achived 92% accuracy on test set,

bigger EfficientNet like B1 or B2 probably will give even better results that this.

All the experiments was documented using Weights&Biases.

[Link to Wandb project](https://wandb.ai/filnow42/skin-cancer)

After experiments I wrote training script using PyTorch Lightning.