VariableNeuralNets: attention-based parameter swapping for continual learning

official submission for the 2022 AWS Deep Learning Challenge

Brain Tumor	Cell Nuclei	Colon Tumor

This project explores the use of attention-based parameter swapping for high accuracy performance across different tasks on the same neural network.

Set-up Steps

0. Set up DL1 Instance / Choose Habana® Deep Learning Base AMI (Ubuntu 20.04) image

1. Select a TensorFlow framework image to run on the above Habana image:

docker pull public.ecr.aws/habanalabs/tensorflow-installer:2.8.0-ubuntu20.04-1.3.0-latest

Run the docker image:

docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host public.ecr.aws/habanalabs/tensorflow-installer:2.8.0-ubuntu20.04-1.3.0-latest

Run the model

Create a working directory for python scripts on EC2 and execute:

• trainUNet.py to train the MainNet (UNet)
• AttentionNet.py to train the AttentionNet (CNN)
• Mechanics.py with given input image to test entire model on given task for either brain tumor, colonoscopy, or cell nuclei segmentation

Use my weights trained on the DL1 instance here.

DL1 / Efficiency Notes

The above models rely on Habana modules to train using AWS DL1 instances. Training across multiple devices results in ~10 epochs under 60 seconds on the AttentionNet (CNN) and under 5 minutes for the MainNet (UNet) using the corresponding datasets referenced below. Both models converge to > 0.85 accuracy.

• For documentation on TensorFlow migration and training see Habana Documentation

Datasets

Nuclei Segmentation: https://www.kaggle.com/c/data-science-bowl-2018/overview

Brain Tumor: https://www.med.upenn.edu/cbica/brats2020/data.html

Colonoscopy: https://paperswithcode.com/dataset/kvasir

---

Submission

Devpost

AWS ID: joshuabelofsky@gmail.com