gt-skin-models

GT-Skin / HAM10000 Skin Lesion Classification

This project is a Google Colab-style PyTorch notebook for multi-class skin lesion classification on the HAM10000 dataset. The workflow in Untitled2.ipynb downloads the dataset, balances the class distribution with augmentation, trains three CNN backbones, and evaluates an ensemble with test-time augmentation.

What the project does

• Loads the skin cancer dataset from Hugging Face Datasets.
• Converts the problem into 7-class classification:
  • nv - melanocytic nevi
  • mel - melanoma
  • bkl - benign keratosis-like lesions
  • bcc - basal cell carcinoma
  • ak - actinic keratoses
  • vasc - vascular lesions
  • df - dermatofibroma
• Resizes images to 299x299.
• Handles class imbalance by generating augmented samples until each class reaches the target size.
• Trains three pretrained backbones:
  • Xception
  • EfficientNet-B5
  • ConvNeXt-Base
• Uses weighted cross-entropy, label smoothing, gradient clipping, and a two-stage training schedule.
• Combines model predictions with an ensemble and test-time augmentation.
• Saves plots and checkpoints for later download.

Repository files

• Untitled2.ipynb - main notebook containing the full pipeline.
• xcp_best.pth - saved Xception checkpoint stored as a large local artifact.
• b5_best.pth - saved EfficientNet-B5 checkpoint stored as a large local artifact.
• cvx_best.pth - saved ConvNeXt-Base checkpoint stored as a large local artifact.
• class_imbalance.png - class distribution plot.
• balance_comparison.png - before/after balancing plot.
• augmentation_demo.png - augmentation examples.
• final_results.png - final evaluation visualization.

Requirements

The notebook installs most dependencies itself, but the core stack is:

• Python 3.10+
• PyTorch
• timm
• albumentations
• datasets
• scikit-learn
• seaborn
• pandas
• matplotlib
• Pillow

How to run

In Google Colab

1. Open Untitled2.ipynb in Colab.
2. Run the dependency installation cell.
3. Run the dataset download and preprocessing cells.
4. Train the three models.
5. Run the ensemble evaluation cell to generate the final metrics and plots.

Locally

The notebook is written with Colab paths such as /content/... and google.colab.files, so it is not fully local-ready without small edits. To run it locally, update the save paths and remove the Colab-specific upload/download cells.

Expected outputs

Running the notebook produces:

• Balanced training/validation/test splits.
• Saved checkpoints for each model.
• Final classification report and confusion matrix.
• PNG plots showing class imbalance, balancing, augmentation, and final results.

Notes

• The notebook is focused on classification only; it does not use segmentation masks.
• The saved .pth files in the workspace are the trained model weights referenced by the notebook.
• If you want, I can also turn this into a cleaner README with usage examples, environment setup, and a project diagram. #� �c�o�n�t�e�x�t��f�r�e�e��s�k�i�n� � �