android13/hardware/rockchip/rknn-toolkit2/examples/functions/mmse/README.md

45 lines
2.4 KiB
Markdown

# How to use MMSE function
## Model Source
The model used in this example come from the following open source projects:
https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md
## Script Usage
*Usage:*
```
python test.py
```
*Description:*
- The default target platform in script is 'rk3566', please modify the 'target_platform' parameter of 'rknn.config' according to the actual platform.
- If connecting board is required, please add the 'target' parameter in 'rknn.init_runtime'.
- The 'quantized_algorithm' parameter of 'rknn.config' is set to 'mmse'. and a 'MmseQuant2' progress bar can be seen during the conversion process, indicating the execution progress of MMSE.
## Expected Results
This example will outputs the results of the accuracy analysis and print the TOP5 labels and corresponding scores of the test image classification results, as follows:
```
layer_name simulator_error
entire single
-----------------------------------------------------------------------------------------------------------
[Input] input:0 1.000000 1.000000
[exDataConvert] input:0_int8 0.999986 0.999986
[Conv] MobilenetV1/MobilenetV1/Conv2d_0/BatchNorm/FusedBatchNorm:0
[Clip] MobilenetV1/MobilenetV1/Conv2d_0/Relu6:0 0.999986 0.999986
...
[Clip] MobilenetV1/MobilenetV1/Conv2d_13_pointwise/Relu6:0 0.858769 0.999334
[Conv] MobilenetV1/Logits/AvgPool_1a/AvgPool:0 0.948201 0.999804
[Conv] MobilenetV1/Logits/Conv2d_1c_1x1/BiasAdd:0 0.963938 0.999562
[Reshape] MobilenetV1/Logits/SpatialSqueeze:0_int8 0.963938 0.999906
[exDataConvert] MobilenetV1/Logits/SpatialSqueeze:0 0.963938 0.999906
```
```
-----TOP 5-----
[155]: 0.9931640625
[154]: 0.00266265869140625
[204]: 0.0019779205322265625
[283]: 0.0009202957153320312
[194]: 0.0001285076141357422
```
- Note: Different platforms, different versions of tools and drivers may have slightly different results.