mnist-train-predict

Train and deploy mxnet gluon models for mnist prediction.

Implentation Details

• Each convolutional layer is followed by batch normalization & relu activation
• Downsampling is achieved through 3 by 3 maxpooling filter with stride 2
• Dropout regularization can be applied between fully connected layers
• Softmax function at output layer
• Cross entropy loss with stochastic gradient descent used to optimize weights

Training

Venv

• Create a virtual env & install requirements. Eg:
  • mkvirtualenv -a ./ -r requirements.txt -p python3 mnist
• Train the model with default hyperparameters (time consuming without gpu):
  • cpu: (mnist) $ python train.py
  • gpu: (mnist) $ python train.py --gpus=1
  • Model parameters will be saved to the checkpoint folder.
• Launch the model server locally:
  • (mnist) $ python server.py --params-file=./checkpoint/epoch9.params
• Download and unzip mnist jpegs
• Send a request to the server
  • curl -X POST -F "image=@<path to a jpg you downloaded>" http://localhost:8080/transform

Notes

• Hyperparameters have not been optimized. This should be done with access to a gpu.
• Model server is cpu only.
• train.py has been factored to allow for easy integration with Amazon Sagemaker hyperparameter optimization.