Using sh cloud.sh will:
- Build the docker image with nvidia support (the base image is supplied by google).
- Upload it to google cloud platform.
- Start a job that can be found in AI-Platform > Jobs.
- Stream the output logs to the terminal session.
The jobs will:
- Take several arguments from command line.
- Run on the GPU if it is avaible.
- Define a sequential CNN.
- Train the model on the dataset (potentially stored on GCP) downloaded from Kaggle.
- Upload the trained model parameters and the evaluation results to google storage bucket.
The scripts are easiliy modified to add features such as reading a pre trained models from google storage to resume a former training session. Ideally specific paths should be specified
This repo is an example for future projects that I cannot reasonably run locally. I followed the google supplied examples closesly, but this is smaller baseline suited to my own one-man-army projects.