https://github.com/zhangge6/netron-hierarchy
View DL models in a hierarchical fashion.
https://github.com/zhangge6/netron-hierarchy
Last synced: 4 months ago
JSON representation
View DL models in a hierarchical fashion.
- Host: GitHub
- URL: https://github.com/zhangge6/netron-hierarchy
- Owner: ZhangGe6
- License: other
- Created: 2024-07-23T14:38:37.000Z (10 months ago)
- Default Branch: master
- Last Pushed: 2025-01-12T14:27:29.000Z (5 months ago)
- Last Synced: 2025-01-12T15:32:42.474Z (5 months ago)
- Language: JavaScript
- Homepage:
- Size: 3.37 MB
- Stars: 7
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
Awesome Lists containing this project
README
# Introduction
[netron](https://github.com/lutzroeder/netron) is great. However, it struggles for models with a large number of nodes. Too much time is consumed for graph rendering, which causes the visualization to get stuck.
Then [`netron-hierarchy`](https://github.com/ZhangGe6/netron-hierarchy) comes to help. It is based on the observation that deep learning models can be viewed in a hierarchical way. More specifically:
- Deep learning models are generally stacks of the same blocks;
- Each block is composed of several high-level layers, like `Attention`, `LayerNorm`;
- In the most fine-grained view, each layer is composed of "atomic" modules, such as `Linear`, and `Add`.
The higher the hierarchy is, the coarser the node grainity is. `netron-hierarchy` enables us to visualize models in a specified hierarchy and switch between hierarchies. As a bonus, hierarchical visualization helps us understand the model topology better.
> When too many nodes are detected in a hierarchy, stacks except the 1st one will be folded automatically, to avoid graph rendering hang.
`netron-hierarchy` is built based on netron. Hope it helps!
# Get started
Clone the repo and launch the application
```bash
git clone https://github.com/ZhangGe6/netron-hierarchy.git
cd netron-hierarchy
python package.py build start --browse
```
Then `netron-hierarchy` will be hosted in `http://localhost:8080` automatically in the web browser.
# Usage
Currently, `netron-hierarchy` parses hierarchy level from node names. So ensure the node names of the provided model have patterns like
```
XXX/block_list.[X]/...
or
XXX/layer.[X]/...
```
Take the [LlamaV2_7B_float16.onnx](https://huggingface.co/alpindale/Llama-2-7b-ONNX/resolve/main/FP16/LlamaV2_7B_float16.onnx?download=true) as example (Yes, LLM models can be opened!), the following images show the model graph of hierarchy 2 and 3, respectively. Hitting "+" or "-" to switch hierarchy levels.
