onnx转ncnn报错：identity not supported yet「用optimizer去掉identity层」

电玩女神 2022-03-26 02:24 727阅读 0赞

**目录**

1、现象描述：identity not supported yet

2、解决方法：使用onnx optimizer去掉identity层

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

### 1、现象描述：identity not supported yet ###

onnx转换ncnn报错：identity not supported yet

onnx可视化（最后多了个identity层）：identity就是f(x)=x。  
![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MTc3MDE2OQ_size_16_color_FFFFFF_t_70][]

据大神解释，转onnx就会出现这种迷之op：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MTc3MDE2OQ_size_16_color_FFFFFF_t_70 1][]

### 2、解决方法：使用onnx optimizer去掉identity层 ###

代码：

from TYY_model import TYY_MobileNet_reg
    from keras.utils.vis_utils import plot_model
    import onnxmltools
    import onnx
    from onnx import optimizer
    
    #读取mobilenet keras
    weight_file = 'mobilenet_reg_0.25_64_asian.h5'
    all_file = 'mobilenet_reg_0.25_64_asian_save.h5'
    img_size = 64
    alpha = 0.25
    base_model = TYY_MobileNet_reg(img_size,alpha)()
    base_model.load_weights(weight_file)
    base_model.save(all_file)
    plot_model(base_model, to_file='model_mobilenet_asian_save.jpg', show_shapes=True)
    
    #keras转onnx
    all_file_onnx = 'mobilenet_save.onnx'
    onnx_model = onnxmltools.convert_keras(base_model)
    onnx.save(onnx_model, all_file_onnx)
    
    #去掉identity层
    all_passes = optimizer.get_available_passes()
    print("Available optimization passes:")
    for p in all_passes:
        print('\t{}'.format(p))
    print()
    
    onnx_optimized = 'mobilenet_optimized.onnx'
    passes = ['eliminate_identity']
    optimized_model = optimizer.optimize(onnx_model, passes)
    onnx.save(optimized_model, onnx_optimized)

打印：

> Available optimization passes:
> 
> eliminate\_identity
> 
> eliminate\_nop\_pad
> 
> eliminate\_nop\_transpose
> 
> eliminate\_unused\_initializer
> 
> extract\_constant\_to\_initializer
> 
> fuse\_add\_bias\_into\_conv
> 
> fuse\_bn\_into\_conv
> 
> fuse\_consecutive\_squeezes
> 
> fuse\_consecutive\_transposes
> 
> fuse\_transpose\_into\_gemm
> 
> lift\_lexical\_references
> 
> nop
> 
> split\_init
> 
> split\_predict

使用字段：「eliminate\_identity」去除identity层

图示：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MTc3MDE2OQ_size_16_color_FFFFFF_t_70 2][]

官网参考信息：

[https://github.com/onnx/onnx/blob/master/docs/PythonAPIOverview.md\#optimizing-an-onnx-model][https_github.com_onnx_onnx_blob_master_docs_PythonAPIOverview.md_optimizing-an-onnx-model]

[https://github.com/onnx/onnx/blob/master/onnx/examples/optimize\_onnx.ipynb][https_github.com_onnx_onnx_blob_master_onnx_examples_optimize_onnx.ipynb]

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MTc3MDE2OQ_size_16_color_FFFFFF_t_70]: /images/20220326/f1dd5aca2d124e9abd93dae3e8505ec1.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MTc3MDE2OQ_size_16_color_FFFFFF_t_70 1]: /images/20220326/942eb409333346ebbe18040b843e01f5.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3dlaXhpbl80MTc3MDE2OQ_size_16_color_FFFFFF_t_70 2]: /images/20220326/698b017e090d412fabb807194832122d.png
[https_github.com_onnx_onnx_blob_master_docs_PythonAPIOverview.md_optimizing-an-onnx-model]: https://github.com/onnx/onnx/blob/master/docs/PythonAPIOverview.md#optimizing-an-onnx-model
[https_github.com_onnx_onnx_blob_master_onnx_examples_optimize_onnx.ipynb]: https://github.com/onnx/onnx/blob/master/onnx/examples/optimize_onnx.ipynb