mldl_htwg

Code accompanying the Machine and Deep Learning Lectures. Last updated: Winter Semester 2025/2026.

Setup for Deep Learning

This repository uses:

• the Data Science Stack (NumPy, Pandas, Matplotlib, etc.)
• Keras 3 with PyTorch and/or JAX as backend
• uv as the package/environment manager (a much faster, modern alternative to plain pip) 🚀🚀🚀

To test the

---

1. Basic Data Science Stack (generic, all platforms)

1.1 Create a virtual environment

We call it htwg_dl_25 and use uv to create it right inside the repository:

uv venv htwg_dl_25

1.2 Activate it

source htwg_dl_25/bin/activate

1.3 Install scientific + Jupyter stack

uv pip install numpy scipy pandas matplotlib seaborn tqdm scikit-learn jupyterlab ipykernel

---

2. Deep Learning Stack (generic)

2.1 Install Keras 3

uv pip install "keras>=3.0"

Keras 3 is backend-agnostic and can run on top of PyTorch or JAX or TensorFlow (not recommended anymore). You select the backend in your code before importing Keras via:

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax"
import keras

---

3. Platform-specific instructions

PyTorch (CPU-only, generic)

uv pip install torch torchvision torchaudio

For CUDA-enabled Linux machines, follow the official instructions on the PyTorch website.

JAX (CPU-only, generic)

For non-Apple-Silicon platforms (or Intel Macs):

uv pip install jax jaxlib

For CUDA-enabled JAX on Linux, follow the official JAX installation guide.

---

About uv

It's worth considering uv as our environment and package manager because it is fast, lightweight, and fully compatible with standard Python workflows. It uses a global package cache, so packages are downloaded and built only once, and subsequent environments simply link to them. This saves both installation time and disk space.

Conceptually, you can think of uv as:

• a drop-in replacement for pip (you just prefix commands with uv)
• faster (Rust-based, with global caching)
• more reliable (modern dependency resolver)
• using standard Python venvs instead of a separate ecosystem like conda

Example:

• pip install numpy → uv pip install numpy
• python -m venv env → uv venv env #create a virtual environment

Everything else works the same — just faster and cleaner.