Training Mnist with Multiple GPU

This is a small script to test the training of a ConvNet model using multi gpu. It is reported that there are some problems when saving the h5 file. It seems that these problems are not present in models like vgg-X but it seems to appear in models with different building blocks (nasnet, resnet, resnext, densenet).

This script also serves as a fast start up to test anaconda environment and use of Hardware Resources (aka GPU!)

Tensorflow requirements:

These scripts have been produced and tested with Tensorflow v1.13.1. Changes may be required for adapting the code to newer versions of tensorflow. Anyway, we provide a yaml file to clone our working environment and fast testing. However, we advice that the environment contains other python libraries and uses at most 5 GB of disk space.

Cloning anaconda environment file

From base environment in anaconda use:

conda env create -f tf_gpu_cuda_100.yaml

Scripts Provided:

• main.py -> This script downloads Mnist

• load_and_test.py -> This script aims to load the saved models from the h5 file and evaluate them to confirm that trained model was sucessfully saved

If this was of use leave a comment and share. Happy coding.