NeuralDBG

When PyTorch training breaks, get ranked causal hypotheses — not another loss curve.

pip install neuraldbg Open in Colab GitHub Public benchmark

Neural Suite

NeuralDBG is the open core. The full stack adds visualization and autonomous remediation (private beta).

Public · PyPI

NeuralDBG

Causal hooks, semantic events, JSON diagnostic export, MIT license.

Private beta

Diagnostic Workspace

Desktop viewer for causal graphs, loss traces, and hypothesis inspection.

Private beta

Neural Agent

Closed loop: diagnose → patch LR, clipping, activations → retry training.

Early access: open a GitHub issue with label suite-access.

Why NeuralDBG?

TensorBoard / W&B

Shows when metrics moved. You still guess the root cause.

NeuralDBG

Explains why with layer-level hypotheses and confidence scores.

Reproducible proof

Run the public benchmark locally — results are checked into the repo.

—

pip install neuraldbg && python -m benchmark_public.run

Quick start

from neuraldbg import NeuralDbg

with NeuralDbg(model) as dbg:
    for step, (x, y) in enumerate(loader):
        dbg.step = step
        loss = train_step(x, y)
        dbg.record_loss(loss.item())

print(dbg.explain_failure())