Tips and Tricks of the Trade

This guide is an ever growing list of issues or setups I personally encounter frequently enough to document here, but unfrequently enough that I forget them otherwise. Feel free to browse!

ZSH Installation

A Power Shell is a huge timesaver. It'll help you navigate through files a lot faster with nice autocomplete, amongst other features.

Quick Install

Install zsh from here
Install the theme here

Brand new ec2 instance for ML research!

In order to do ML research, choose a g5.large with at least 250 GB of storage to load all the modules and models. It's not too expensive - something like $300 a year. It'll be quite expensive, so remember to stop the machine

I prefer to start with an ubuntu machine as the aws instance comes with CentOS I believe. The aws instance has performance improvements, but for jupyter notebook work it shouldn't matter.

Once you're set up, ssh into the machine and let's get started!

First things first, if you're gonna use a notebook, this will save you a lot of trouble:

echo "alias start='source myenv/bin/activate" >> ~/.bashrc
echo "alias jupyter='jupyter notebook --no-browser --port 1234'" >> ~/.bashrc
source ~/.bashrc

Get your machine up to date on the latest packages:

sudo apt update
sudo apt upgrade -y
sudo apt install git 
sudo apt install gh
gh auth login

It'll ask you to reboot the machine. Go do it.

Install Python. This should install the latest distribution.

sudo apt install python3 python3-pip build-essential libssl-dev libffi-dev python3-dev pkg-config

Install a virtual environment (you can see below for more details)

Note that pip3 and pip are the same :)

sudo apt install python3.10-venv
pip3 install virtualenv
python3 -m venv myenv
source myenv/bin/activate

And voila - you should be in the env. Now come the packages...

Install all the pip menagerie

pip install numpy pandas scipy matplotlib scikit-learn tensorflow jupyter awscli jupyterlab-vim boto3 python-dotenv tqdm openai opencv-python Pillow diffusers transformers accelerate safetensors pytz boto3 rsa python-dotenv omegaconf mediapipe opencv-python-headless controlnet-aux xformers PySocks
aws configure
python -m ipykernel install --user --name=myenv --display-name="MyProject"

And lastly, some cuda stuffs! First, check if you have GPUs with lspci. If you do have an NVIDIA GPU (mine shows like this: 00:1e.0 3D controller: NVIDIA Corporation GA102GL [A10G] (rev a1)), do the below, which I got from here

sudo apt install nvidia-cuda-toolkit
sudo apt-get upgrade -y linux-aws
sudo apt-get install -y gcc make linux-headers-$(uname -r)
cat << EOF | sudo tee --append /etc/modprobe.d/blacklist.conf
GRUB_CMDLINE_LINUX="rdblacklist=nouveau"
sudo update-grub
aws s3 cp --recursive s3://ec2-linux-nvidia-drivers/latest/ .
chmod +x NVIDIA-Linux-x86_64*.run
sudo /bin/sh ./NVIDIA-Linux-x86_64*.run
sudo touch /etc/modprobe.d/nvidia.conf
echo "options nvidia NVreg_EnableGpuFirmware=0" | sudo tee --append /etc/modprobe.d/nvidia.conf

nvidia-smi -q | head

And that's it! Now go ahead and fully reboot the instance...

sudo reboot

Virtual Environments

If you want to work with a virtual env (which you absolutely should if you're not using a containerized solution), run these commands:

NEWENVNAME=tryiton
NEWENV_PYTHON_VERSION=python3.11

pip3 install virtualenv
virtualenv --python=$NEWENV_PYTHON_VERSION $NEWENVNAME
source $NEWENVNAME/bin/activate
...
python3 -m pip install ipykernel
python -m ipykernel install --user --name $NEWENVNAME --display-name "$NEWENVNAME"
...
deactivate

Jupyter and Virtual Envs

Sometimes the jupyter notebook kernel doesn't start where it should be (sys.executable returns the wrong python kernel). This is because our kernel list isn't updated properly. To do so, go to ~/.local/share/jupyter/kernels/{envname}/kernel.json and change the first parameter in the argv list to the python path in your specific kernel.

{
 "argv": [
  "path/to/kernel/python",
  "-m",
  "ipykernel_launcher",
  "-f",
  "{connection_file}"
 ],
 "display_name": "Python 3 (ipykernel)",
 "language": "python",
 "metadata": {
  "debugger": true
 }

SSH Keys

Guide for this.

Generate the SSH key you want to use with the following command:

$ ssh-keygen -t rsa -C "your_name@email_domain.com"

Make sure the key is cached by typing:

$ ssh-add -l

If no keys appear, then add the key by typing:

$ ssh-add ~/.ssh/id_rsa

To have all connections linked to the same id_rsa key, add this line of code to ~/.ssh/config:

Host *
    AddKeysToAgent yes
    PreferredAuthentications publickey
    IdentityFile ~/.ssh/id_rsa

Lastly, check that the connections work:

$ ssh -T git@github.com
Hi {user}! You've successfully authenticated, but GiHub does not provide shell access.
$ ssh -T git@gitlab.work.io
Welcome to GitLab, @{user}!

Viewing Jupyter Notebooks Locally from a Remote Machine

Jupyter notebook development is becoming the de-facto research environment. To work with Jupyter notebooks remotely, one must tunnel the output of jupyter to view it on a local browser. You can also use a reverse proxy, but I find this more troublesome.

ssh username@xx.xx.xx.xx -L 1234:localhost:1234

In the code above, data outgoing to port 1234 will be forwarded to localhost:1234 from the remote system. The L is used to specify the port to forward data to.

Once you're in the machine, start the jupyter notebook as such:

jupyter notebook --no-browser --port 1234

This will forward the output of the jupyter notebook to the port specified.

landmann/tips-and-tricks