Cats vs Dogs

Motivation

How to train a deep learning model when I don't have enough hardware power to train the model, and even store all the tons of data that the model requires?

Well... I go to the cloud.

Objective

This repository examplifies how to traing a job on GCP when the input data is too big to be able to store it locally; in other words, the input data is also in the cloud, more specifically in a GCS bucket.

The model itself consist of a Convolutional Neural Network, build and trained purely on TensorFlow, in order to classify images from cats and dogs. The objective of this repository it's not to enter into the details of deep classifier models, but to show how to encapsulate a working package and send a training job to the Google Compute Engine.

Input Data

As an input data I used the well known Kaggle Cats Vs. Dogs dataset. Please update the dataset into your Cloud Storage Bucket following this hierarchy of folders:

|- <BUCKET_NAME>
      └── cats_vs_dogs
          └── trained_models
          └── data
              └── train
                  └── cats
                  └── dogs
              └── validation
                  └── cats
                  └── dogs

Where the trainining and validation JPEG images goes inside the respective cats or dogs folder.

Training on the Cloud

After making sure that you have installed the lastest version of the Google Cloud SDK, just run the following command:

source train_cloudml.sh <BUCKET_NAME>