├── .gitignore
└── readme.md


/.gitignore:
--------------------------------------------------------------------------------
1 | .vscode/
2 | .vscode/
3 | 


--------------------------------------------------------------------------------
/readme.md:
--------------------------------------------------------------------------------
 1 | # WNNX Models
 2 | 
 3 | wnnx is a new model format which exports models directly from pytorch without onnx. It can be infered via **wnn** framework, the infer lib is also incrediablly simple and insamely small to serveral kbs.
 4 | 
 5 | Even though, **wnn** can achieve a extremly fast speed and low memory consumption in both X86 and arm platform.
 6 | 
 7 | **wnnx_models** contains some popular models which exports from pytorch. It can be viewed using:
 8 | 
 9 | ```
10 | pip install wnetron
11 | wnetron .
12 | ```
13 | 
14 | When wnnx merged by `netron` official repo, you can also view *wnnx* model with `netron` as well.
15 | 
16 | 
17 | ## Results
18 | 
19 | - `SimpleFC`:
20 | 
21 | ![](https://raw.githubusercontent.com/jinfagang/public_images/master/20220622155647.png)
22 | 
23 | - `MobilenetV3`:
24 | 
25 | ![](https://raw.githubusercontent.com/jinfagang/public_images/master/20220622155829.png)
26 | 
27 | - `DynamicLayernorm`:
28 | 
29 | ![](https://raw.githubusercontent.com/jinfagang/public_images/master/20220622160247.png)
30 | 
31 | 
32 | As you can see, wnnx can be fully replace onnx as an option to inference model. It has many strength compare with onnx:
33 | 
34 | - Powered with flatbuffer, instead using protobuf which is heavy;
35 | - Easy to convert from trained model via pytorch;
36 | - Without **any** glu ops, the whole graph is extremly clean;
37 | - Can be efficiently inferenced via **wnn**.
38 | 
39 | 
40 | 


--------------------------------------------------------------------------------