Information
This image will train a VGG16 model for the imagenette dataset. Almost all the information out there is for using a pretrained model or training on a very simple image dataset. This image attempts to create a recreate-able way to generate the weights for the full model in keras.
How to build
The following command will build the image:
docker build -t vgg_training:16 -f Dockerfile .This will get the .tgz data and put it into the image for later use.
How to run
By default this trains on CPU. To train with GPU, its advised to use the nvidia-docker runtime. Once that is installed, you can run something such as: sudo docker run --rm --gpus all vgg_training:gpu
To use that follow the instructions at: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html
To run simply use docker run vgg_training:16
What is expected
A trained model is trained in the image. To save the model, mount the output folder in the image to somewhere, for instance: docker run -v $(pwd):/root/output vgg_training:16 or set the ENV OUTPUT on run.