forkpoint

Visual model surgery for CNNs.

Load a PyTorch model. See its weights as visual mass. Click a layer, ablate it, fork the model, run inference on both, and diff what changed: in the prediction and on the graph itself.

---

Demo

TODO: drop a 10 to 15 second GIF here showing load → click layer → ablate → fork → diff. docs/demo.gif

---

What it does

Step	Action
Load	Point it at a .pt / .pth checkpoint or a torchvision model name.
See	The graph renders with weight magnitude mapped to visual mass, so heavy layers look heavy.
Cut	Click a layer to zero it. No config file, no re-instantiation.
Fork	The edited model branches off the original. Both stay loaded.
Diff	Run the same input through both and compare outputs side by side, with changed regions highlighted on the graph.

---

Quickstart

TODO: verify these against a clean clone before publishing.

# backend
cd server
uv venv && uv pip install -r requirements.txt
uvicorn app.main:app --reload

# frontend
cd web
npm install
npm run dev

Open http://localhost:5173.

---

Stack

Backend: Python 3.14, PyTorch 2.12, FastAPI Frontend: Vite, React, TypeScript, Tailwind, d3 Models: CNNs first (torchvision ResNet, YOLO)

---

Why this exists

Nothing currently does load model → see it as visual mass → click → ablate → fork → re-run → diff.

Tool	Interactive edits	GUI	Compares model variants
Netron	no	yes	no
nnsight	yes	no (notebook library)	partial
W&B / TensorBoard	no	yes	training runs, not variants
forkpoint	yes	yes	yes

The gap is real. This fills it.

---

Non-goals

• Not a training tool. Inference only.
• Not a deployment tool. No ONNX export, no quantization.
• Not a TensorBoard replacement. Those compare training runs over time; this compares model variants live.
• Not architecture-general. Transformers are out of scope for v0.

---

Roadmap

Version	Scope
v0	Layer-zeroing as the single surgery primitive. CNNs only.
v0.1	Per-channel and per-filter ablation.
v0.2	Activation diffing on intermediate layers.
v0.3	Real knowledge distillation between forked variants.

---

Status

Pre-alpha. Built incrementally, APIs change without notice, nothing is stable yet.

---

_{forkpoint is a working name. The repo directory is still KD tool for historical reasons.}