keras_cnn_fashion_mnist

Shallow CNN for Fashion MNIST using Keras and SageMaker

Overview

This project contains code and notebooks for training a custom CNN on the Fashion MNIST dataset either locally or in the cloud using Amazon SageMaker.

The CNN is written with Keras and Tensorflow backend. Generic and Fashion MNIST specific-versions are implemented as classes in cnn.py.

Directories

• data/ - Fashion MNIST data files
• models/keras_checkpoints - Keras checkpoints

Training scripts

To train the network on your local machine:

python train_script_local.py

To train in the cloud using Amazon SageMaker use train_script_sagemaker.py and script mode. This can be done locally or it can be done in a Sagemaker notebook instance ¹.

Notebooks

There are two Jupyter notebooks:

• explore_data_and_model.ipynb -- An introductio to the dataset and default model
• train_tune_test.ipynb - training, tuning and testing the model using Sagemaker resources.

Environments.

If you're running the scripts or notebooks locally, it's recommended to create a virtual environment directly from the included environment files

Using virtualenv

python3 -m venv fashion
source env/bin/activate
pip install -r requirements.txt

Using conda

conda env create -f environment.yml

If you're running anything in a SageMaker notebook instance, you can use the built-in conda_python3 kernel, provided you install keras ²

Footnotes

1. The easiest way to get up and running in a SageMaker notebook instance is probably to fork this repo and link it to the notebook instance. ↩

2. At the time of writing, this wouldn't work using a Lifecycle configuration due to a timeout, but you can install directly from a Python notebook within the instance using ! conda install --name conda_python3 keras ↩