search

rosinusserrano / pml_vqvae

Repository for the course "Project Machine Learning" during WiSe 24/25 at TU Berlin consisting of a replication of the paper "Neural Discrete Representation Learning" (van den Oord et al., 2018).
0 stars 0 forks source link

readme

pml_vqvae

How to run

Using the CLI

To run start the training, you can run the training script with python src/pml_vqvae/train.py

On default (without extra parameters givin in the command), the training script will look up the config.yaml-file and will train according those defined parameters.

If you want to change the default values you can either change them in the config.yaml or you can pass the values yu want to change via the cli, e.g.

python src/pml_vqvae/train.py --learning_rate 0.1

This will take the config.yaml as the foundation and overwrites all the given values passed in the command. See the python src/pml_vqvae/train.py -h for more information.

Running on Cluster

  1. ssh to cluster entry-node by ssh <username>@hydra.ml.tu-berlin.de
  2. Build container if not exist, see Section Build Container
  3. Request a GPU node srun --partition=gpu-test --gpus=1 --pty bash
  4. Go to project directory cd pml_vqvae
  5. Run ./run_script.sh

Build Container

  1. Obtain a node on the cpu only partition with enough run time srun --partition=cpu-2h --pty bash
  2. Switch to the project root (pml_vqvae)
  3. Build the container with apptainer build pml.sif pml.def (can take up to 15min)

How to Log with Weights and Biases

  1. You need to have a Weights and Biases (wandb) account
  2. Add your wandb api key to your environment variables export WANDB_API_KEY=<your key>. If you run train the model within an apptainer make sure to add the environment variable to your pml.def-file
  3. Create an experiment by adjusting the config.yaml-file. make sure log_wandb=True

When editing LaTex in VS Code

Setup

  1. Install TexLive (best would be to install the full version)
  2. Install Latex Workshop VS Code extension
  3. Make sure you have pulled the git version where there is a .vscode/settings.json file containing settings for latex workshop and a gitignore that has the file endings of auxiliary latex build files

Workflow

TL;DR

Open the .tex file you want to edit, press Ctrl + Alt + B (Windows/Linux), Cmd + Option + B (macOS) to build the PDF and then Ctrl + Alt + V (Windows/Linux), Cmd + Option + V (macOS) to view the PDF. Then edit your file and you should get live updates in the PDF view.

Shortcuts

Generally, you should be able to open the .tex document that you are interested in and use the following commands to compile, view, and navigate the document:

  1. Compile the LaTeX document

    • Shortcut: Ctrl + Alt + B (Windows/Linux), Cmd + Option + B (macOS)
    • Description: Compiles the main .tex file using the default recipe (e.g., latexmk).

  2. View PDF Preview

    • Shortcut: Ctrl + Alt + V (Windows/Linux), Cmd + Option + V (macOS)
    • Description: Opens a PDF preview window beside the editor to view the compiled PDF. The preview updates automatically after each successful compilation.

  3. Sync PDF with Source (Forward Sync)

    • Shortcut: Ctrl + Alt + J (Windows/Linux), Cmd + Option + J (macOS)
    • Description: Jumps from the .tex source in the editor to the corresponding location in the PDF preview.

  4. Sync Source with PDF (Reverse Sync)

    • Shortcut: Ctrl + Click in the PDF Preview
    • Description: Jumps from a location in the PDF preview back to the corresponding place in the .tex source