[k210 stepping pit] pytorch model is converted to kmodel and used on dock. (ultimately not achieved)

Preface

Hardware ：maix dock
Software ：maix ide nncase 0.1（ Version big pit ）

One 、 Step on the big pit

1. maix Yes k210 Of nncase Model transformation tools have long ceased to support , Only support nncase 0.1 edition . Use the latest nncase The conversion kmodel stay maix Can't load , Take another one. , So we need to use maix Yes k210 Do development , To use nncase0.1 To transform the model , But the official website of Jianan has already been updated nncase1.7, and 0.1 It belongs to fossil version , No longer do support .【 good , I want to use maix, Only choose nncase0.1】
2. Connect , therefore 0.1 very low,nncase0.1 Conversion only tflite Model to kmodel. And I found in use ,nncase0.1 It seems that it can't be changed pytorch From the model transformation tflite（ Personal understanding ）, I don't know what the principle is , In short, it is to report an error . Switch to a tensorflow Model generated tfliite You can use it nncase0.1 Conversion success .【 good , Then I'll replace it with tensorflow Write model 】
3. Connect ,tensorflow Generated tflite, the nncase0.1 To kmodel, Burn to maix dock in . One problem I found was ,maix Provided model inference function , Input can only be pictures （ Personal understanding ）, But what I want is that the network input data and output data are determined by myself .【 good , That's not necessary maix Developed , Change to the one officially provided by Jianan K210 SDK library 】
4. Connect , Switch to a k210 SDK library , Um. , The whole thing can't come , It's too hard , Official from Jia Nan k210 sdk Routines are for their own development boards . But now I use maix dock Development board , I want to use other development board routines to develop my own development board , ok , The whole thing can't come , pure c Write various drivers and models to load reasoning code .【 good , I give up maix dock k210】
5. As mentioned above , Hit the wall step by step , Change your route step by step , Finally give up 210（ Personal dishes ）, Ready to change v831. The following is a record of the process of this exploration
6. summary ： His writing pytroch The model wants to k210 Upper Department , We have to choose the official development board of Jianan , With what he offered SDK Development .maix dock no way （ Personal understanding ）

Two 、pytorch turn kmodel

The first ① Step ：pytorch Model structures, 、 Training 、 preservation

The first ② Step （ Train of thought ）：pytorch turn onnx turn pb turn tflite turn kmodel

1. pytorch turn onnx （ ripe ,pytorch Official support ）

2. onnx turn pb（）

3. pb turn tflite（）

4. tflite turn kmodel（nncase0.1）

The first ② Step （ Train of thought two ）：pytorch turn onnx turn h5 turn tflite turn kmodel

The first ② Step （ Train of thought three ）：k210 Jianan SDK Development