ai2ys

TensorFlow - Best Practices

Randomization

Understanding Randomization in TF Datasets (including TFRecords) https://colab.research.google.com/github/christianmerkwirth/colabs/blob/master/Understanding_Randomization_in_TF_Datasets.ipynb

Transfer learning/fine-tuning

How to apply transfer learning and fine-tuning. https://www.tensorflow.org/guide/keras/transfer_learning

ai2ys

Docker Context

Create a Docker context for remote Machine (access via SSH).

Create Docker context

# specifiy remote user, remote IP address, context name, and the description
remote_user=<user name>
remote_ip_address=<ip address of remote system>
context_name=<name for the context>

ai2ys

Install docker-compose on Ubuntu 20.04

I tried different installation methods via

1. curl which failed because of a non existing URL
2. apt-get which worked but the installed version does not GPU access
3. pip which worked and the installed version does GPU access

Below the tried methods in detail.

❌ Installing via curl - failed because of URL (404)

Following the instructions from https://docs.docker.com/compose/install/ resulted in the following error when trying to run docker-compose afterwards.

ai2ys

Pre-Built TensorFlow Docker Images Supporting CUDA 11

Newer NVIDA GPUs like the GeForce RTX 3090 require CUDA 11.

Pre-build Docker images provided by TensorFlow

TensorFlow provides pre-built Docker images on dockerhub - tensorflow/tensorflow. The underlying Docker files are availble in the TensorFlow GitHub repository.

Currently latest TF version is 2.7. In table gives an overview of CUDA version used within the pre-built Docker images for the specified TF versions. Versions were retrieved from the TensorFlow GitHub repository (date 2022-01-10).

ai2ys

Specify Package Version apt-get install

Installing a specific version

Install exactly the specified version, example:

# Installing exactly version '1.20.3-1ubuntu2' of wget
sudo apt-get install wget=1.20.3-1ubuntu2

ai2ys

ai2ys

Moving the Docker data directory - Ubuntu 20.04

Resources

• https://www.guguweb.com/2019/02/07/how-to-move-docker-data-directory-to-another-location-on-ubuntu/
• https://docs.docker.com/engine/reference/commandline/dockerd/#daemon-configuration-file

# work as root
sudo su

ai2ys

Using nvidia-smi

Commands on local system

Running nvidia-smi every n seconds

nvidia-smi -l <n seconds>

Retrieving GPU index and name

ai2ys

Managing SSH keys for Github and Gitlab

NOTICE: This guide will help you set ssh keys for GitHub and GitLab. However, this is not going to change your commit user.name or user.email. If you need to change those for specific repositories, just run the following commands while in your repository:

git config user.name "Your Name Here"
git config user.email [email protected]

For more info, see this answer. Also, keep in mind this only changes the .git folder inside your repository which never gets added/committed/pushed/uploaded.

I recently had to manage two ssh keys (one for Github and one for Gitlab). I did some research to find the best solution. I am justing putting the pieces together here.