Adaptive Mini-batching for Faster Convergence over complex data

This is the implementation of AdaMB training Tensorflow

Prerequisites

Following the instructions for Tensorflow on Docker

• Have a GPU and an NVIDIA driver of >384.10
• Install Docker CE
• Install nvidia-docker2

Run training

Using the Dockerfile provided with this repo, build the custon Docker container: docker build -f Dockerfile -t tf_models .

Now using docker, run train.py over a container of the image you just built with the current directory mounted, for example: docker run --runtime=nvidia -u $(id -u):$(id -g) -v $(pwd):/ada_mb -it tf_models python /ada_mb/train.py

Caution: Copy and pasting this direct may sometimes fail depending on your screen as Github or your text editor may insert a neline character.