neuralveil 1.0.0

NeuralVeil CLI — part of the NeuralVeil Tensor Shape Debugger. Graph any PyTorch model architecture in one command using torch.fx and forward hooks. Captures real tensor shapes locally — no server, no code upload, works in air-gapped environments.

neuralveil

Graph your PyTorch model architecture in one command.

pip install neuralveil
neuralveil parse model.py

Drop the output into neuralveil.dev and see your model as an interactive graph — layers, tensor shapes, parameter counts — all verified by actually running your code.

---

How it works

your model.py  ──▶  neuralveil parse  ──▶  your local PyTorch
                                                    │
                                         torch.fx trace / forward hooks
                                                    │
                          neuralveil_output.json  ◀─┘
                                    │
                        drop into neuralveil.dev  ──▶  interactive graph

The CLI tries two capture strategies, in order:

1. torch.fx symbolic trace (preferred) PyTorch's official graph capture API. Traces forward() into a structured FX graph IR with typed nodes, operator types, and shapes. Handles VGG, ResNet, DenseNet, and most CNN families cleanly.

2. register_forward_hook (fallback) For models with dynamic control flow that torch.fx can't symbolically trace. Runs a dummy forward pass with hooks on every named submodule, capturing real input/output shapes and parameter counts. Works on virtually any valid PyTorch model.

---

Installation

pip install neuralveil

PyTorch is listed as an optional dependency — ML engineers already have it. If you're setting up a fresh environment:

pip install "neuralveil[torch]"

Requirements: Python ≥ 3.9

---

Usage

# basic
neuralveil parse model.py

# specify input shape
neuralveil parse model.py --input 1,3,224,224

# custom output path
neuralveil parse model.py --output my_graph.json

If the file defines multiple nn.Module subclasses, the CLI will prompt you to pick one.

Python API

import neuralveil

graph = neuralveil.parse_model("model.py", input_shape=(1, 3, 224, 224))
print(graph["nodes"])

---

Security model

The CLI runs entirely on your machine. There is no server, no code upload, no API key, no rate limit.

• Your code never leaves your environment
• Works fully offline and in air-gapped environments
• Execution timeout: 15 seconds (SIGKILL after)

---

Contributing

Issues and PRs welcome. Built by Gyan Shresth · Apache 2.0