دفاتر الملاحظات والبرامج والبذور القابلة لإعادة الإنتاج

You will write code in notebooks for quick experiments and in plain scripts for repeatable runs. Each suits a different moment.

A notebook runs code in cells, showing plots and results inline. It is perfect for poking at data and trying ideas fast. 📓

A script runs top to bottom in one go. Use it for training jobs you want to launch, share, and reproduce exactly.

Notebook cells keep state even when you run them out of order, which can hide bugs. A script always starts from a clean slate.

Deep learning uses random numbers for weights and shuffling. Run twice and you get different results unless you control that randomness.

A seed fixes the starting point of the random number generator, so the same code produces the same numbers every single run.

Set PyTorch's generator with one call. manual_seed makes weight initialization and any torch randomness repeatable.

Your code often pulls randomness from NumPy and Python too, so seed all three to make a run fully reproducible.

When you train on CUDA, also seed the GPU generators. cuda.manual_seed_all covers every GPU device in your machine.

Bundle every seed call into one set_seed helper. Call it at the top of each run so you never forget a source of randomness.

Reproducible runs let you compare experiments fairly. If accuracy improves, you know it was your change, not random luck.

You saw when to use notebooks versus scripts and how seeding torch, NumPy, and Python makes every run reproducible. 🌱