alexandraabbas / mole-image-classifier

Tensorflow 2.0 & Keras classifier trained on Skin Cancer MNIST: HAM10000 data from Kaggle

Tensorflow 🧡 Keras mole image classifier

Trains sequential Keras model on Skin Cancer MNIST: HAM10000 dataset from Kaggle. Training supports local machine and AI Platfrom.

📒 Contents

• notebooks: Exploratory Jupyter notebooks
• scripts: Bash scripts for local and cloud training
• trainer: AI platform trainer module
  • model.py: Defines a Keras model and input function for training
  • task.py: Trains a Keras model to predict cancer class
  • util.py: Utilities to download and preprocess data

📈 Solution diagram

🚀 Get started

First, install the requirements.

pip install -r requirements.txt

To start training either locally or on AI Platfro you need to create an environment variable called KAGGLE_KEY which has to point to a json file containing your Kaggle API key. For more information about how to generate and download your API key from Kaggle, visit the Kaggle API Github page.

export KAGGLE_KEY="path/to/my/kaggle.json"

💻 Local training

To train the Keras model locally, run train-local.sh.

sh scripts/train-local.sh

If you'd like to update the training parameters, edit train-local.sh file and add your desired values.

☁️ Train on AI Platform

To train the Keras model on AI Platform, you need to create a Google Cloud project. For more information on how to create a project on Google Cloud Platform, visit the documentation.

Next, create a Storage bucket where you can store all the logs, checkpoints and SavedModels. Set the name of the bucket as environment variable.

export BUCKET_NAME="my-bukcet-name"

Enable the AI Platform API in your projects. After that you are all set to run train.sh.

sh scripts/train.sh