comsm0045-applied-deep-learning/tensorflow-labsheets

Lab worksheets for the Applied Deep Learning Course.

README

Lab sheets for the Advanced Deep Learning course

Overview

Labsheet	Description
0	Introduction to Python and the scientific python ecosystem
1	Your First Fully Connected Network
2	Convolutional Neural Networks and Tensorflow Graphs
3	Improving your training through batches, learning-rate and initialisation
4	Data Augmentation
5	Adversarial Training

Problems

Kindly file an issue with a description of the problem you're facing, your setup, what you are observing and what you expect to happen instead.

tensorflow-labsheets's Issues

Add intro to batch jobs

• squeue
• sbatch

Start by expanding the intro in lab 1

All labs: Cache dataset locally

The CIFAR10 server is very unreliable and frequently returns 404 or 500 errors, we need to be able to run lab sessions when the server is down. Let's host the dataset locally on BC4 in some shared storage and use that.

Add Windows FAQ

How to...

• SSH using putty
• scp files (see cmder for a good terminal emulator bundled with openssh
• Do SSH tunnels (again, use cmder)
• Replace CRLF (Windows) line endings with LF, in case they still clone on Windows and upload to BC

Lab 2: Tensorboard must be run from a GPU node

The login node the students are using don't have a working version of tensorflow installed so they need to run tensorboard in the background instead.

Lab 1: Cloning on BC4

Instructions should clone on BC4 rather than locally and scping to avoid \n\r issues prevalent on windows.

Add instructions for running batch jobs

Expand on https://github.com/COMSM0018-Applied-Deep-Learning/labsheets/blob/master/Lab_1_intro/Lab%20sheet%201.ipynb

Fix portfolio folder names

In Part 7 the folders have the incorrect names - specifically the learning rates.

Link to GPU stress test is broken

In Lab Sheet 1.ipynb, the 'GPU Stress Test' link is broken.

I think it should be this: https://github.com/COMSM0018-Applied-Deep-Learning/labsheets/blob/master/Lab_1_intro/Running%20TensorFlow.ipynb

Need BC4 account code for batch jobs

BC4 is set up with several "accounts". The ACRC have added one for this course which is comsm0018. If this is added to the job submission script the jobs will get a better priority in the queue.

Ask people to use .ssh/config

• Set up aliases in .ssh/config
• Create ssh id_rsa if not present
• ssh-copy-id to copy id file across so no need to use passwords

Lab 1: Remove ampersand/background from tensorboard invocation

This seems to confuse people, just get them to open two SSH terminals and do tensorboard in one and an interactive session in another.