What this is

paper2code is a collection of AI/ML research papers rebuilt in Python — stripped of the abstractions that hide what's actually happening.

Every paper gets replicated twice (or close to it):

• 🧱 From scratch — raw NumPy / pure PyTorch tensors. No nn.Module magic, no torch.optim, no black boxes. You read the math, you write the math.
• 📦 With libraries — the canonical high-level implementation, so you can diff the two and see exactly what the library is doing for you.

The goal isn't benchmark-chasing. It's understanding — by the time you've written backprop by hand once, loss.backward() stops feeling like magic.

📚 Paper hitlist

Paper	From scratch	With library	Tests	Folder
Attention Is All You Need (Vaswani 2017)	🟢	🟢	10	attention/
An Image is Worth 16x16 Words — ViT (Dosovitskiy 2020)	🟢	🟢	7	vit/
Adam: A Method for Stochastic Optimization (Kingma 2014)	🟢	🟢	6	adam/
Long Short-Term Memory (Hochreiter 1997)	🟢	🟢	12	lstm/
RNN Encoder–Decoder / GRU (Cho 2014)	🟢	🟢	8	gru/
Faster R-CNN (Ren 2015) — key atoms	🟡	🟢	14	faster-rcnn/
SSD: Single Shot MultiBox Detector (Liu 2016) — key atoms	🟡	🟢	11	ssd/
Bits & pieces from scratch	🟢	—	—	things-from-scratch/

_{🟢 done · 🟡 partial (detection papers — RPN/default-boxes/NMS/HNM from scratch; full pipeline via torchvision) · 🔴 todo}

_{Total: 68 tests, ~1.5s to run the whole repo suite.}

🚀 Get started

git clone https://github.com/hegdeadithyak/PaperReplica.git
cd PaperReplica
pip install -r requirements.txt

# run every test across every paper
python3 -m pytest -v

Each paper folder is self-contained: {name}_scratch.py + {name}_library.py + test_{name}.py + a README with the math, a diagram, and first-principles explanation. cd in and read the README.

🧱 The pattern

Every paper lives in a folder with the same four files:

<paper-dir>/
  <name>_scratch.py    # raw PyTorch tensor ops — no nn.Module, no autograd for RNNs
  <name>_library.py    # torch.nn / torchvision / torch.optim equivalent
  test_<name>.py       # shared-weight parity tests — scratch == library to ~1e-5
  README.md            # math, first-principles walkthrough, diagram from Wikimedia

Scratch impls expose their weights in the same layout as the library module, so tests can copy weights across with load_from_torch_*() and verify bit-identical outputs. No hand-waving, no "similar behavior" — numerical equivalence or the test fails.

📖 References

• Vaswani et al. — Attention Is All You Need
• Dosovitskiy et al. — An Image is Worth 16x16 Words
• Kingma & Ba — Adam: A Method for Stochastic Optimization
• Hochreiter & Schmidhuber — Long Short-Term Memory
• Cho et al. — Learning Phrase Representations using RNN Encoder-Decoder
• Ren et al. — Faster R-CNN
• Liu et al. — SSD: Single Shot MultiBox Detector

🪪 License

MIT — see LICENSE.

_{built for the people who'd rather read the paper than the docs}