The MNIST handwritten digits classification problem is one of the most commonly used benchmarks for introducing machine learning concepts. We present here an implementation of the mnist problem using tensorflow with a couple of different models.
We implement it first using plain tensorflow and python which you can find in the tensorflow directory. Secondly, we implement the same code on the C3 platform in the C3 directory. We use the KerasPipe Type to encapsulate the tensorflow models, and create an IDXFile type to read in the idx files containin the mnist training and test data.
To execute the python implementation, first use conda to install a new environment with the packages contained in env.yaml. (Use env-osx.yaml If you're a mac user.)
conda create -p ./venv -f ./env.yaml
Then, load the environment, and launch jupyter notebook.
conda activate ./venv
jupyter notebook
Finally, open MNIST_Example.ipynb and execute its cells!
To execute the C3 implementation, first provision the code in the directory c3/dtitraining to your C3.ai tag. Once finished, open the jupyter notebook c3/mnistExample.ipynb and connect to your c3 session if necessary. Then execute the notebook cells!