Introduction

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 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. When too many nodes are detected in a hierarchy, stacks except the 1st one will be folded automatically, to avoid graph rendering hang. As a bonus, hierarchical visualization helps us understand the model topology better.

netron-hierarchy is built based on netron. Hope it helps!

Get started

Clone the repo and launch the application

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 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.