39 lines
2.6 KiB
Markdown
39 lines
2.6 KiB
Markdown
|
# How to use hybrid-quantization function
|
||
|
|
|
||
|
|
## Model Source
|
||
|
|
The model used in this example come from:
|
||
|
|
https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf1_detection_zoo.md ssd_mobilenet_v2_coco
|
||
|
|
|
||
|
|
## Script Usage
|
||
|
|
*Usage:*
|
||
|
|
```
|
||
|
|
1. python step1.py
|
||
|
|
2. modify ssd_mobilenet_v2.quantization.cfg according to the prompt of step1.py
|
||
|
|
3. python step2.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'.
|
||
|
|
|
||
|
|
## Expected Results
|
||
|
|
This example will outputs the results of the accuracy analysis and save the result of object detection to the 'result.jpg', as follows:
|
||
|
|
```
|
||
|
|
layer_name simulator_error
|
||
|
|
entire single
|
||
|
|
--------------------------------------------------------------------------------------------------------------------------------------
|
||
|
|
[Input] FeatureExtractor/MobilenetV2/MobilenetV2/input:0 1.000000 1.000000
|
||
|
|
[exDataConvert] FeatureExtractor/MobilenetV2/MobilenetV2/input:0_int8 0.999964 0.999964
|
||
|
|
[Conv] Conv__343:0
|
||
|
|
[Clip] FeatureExtractor/MobilenetV2/Conv/Relu6:0 0.999976 0.999976
|
||
|
|
|
||
|
|
...
|
||
|
|
|
||
|
|
[Conv] BoxPredictor_0/BoxEncodingPredictor/BiasAdd:0 0.997577 0.999687
|
||
|
|
[Transpose] BoxPredictor_0/BoxEncodingPredictor/BiasAdd__341:0 0.997577 0.999960
|
||
|
|
[Reshape] concat_swap_concat_reshape_i0_out 0.997577 0.999960
|
||
|
|
[Concat] concat_swap_concat_reshape_o0_out 0.997513 0.999939
|
||
|
|
[Reshape] concat:0_int8 0.997513 0.999943
|
||
|
|
[exDataConvert] concat:0 0.997513 0.999943
|
||
|
|
```
|
||
|
|
|
||
|
|
- Note: Different platforms, different versions of tools and drivers may have slightly different results.
|