Local model training

Local model training is performed using sipeed/maix_train this code, using Tensorflow as the training framework

Main support:

  • Object classification model (using Mobilenet V1): Only identify what is the object in the picture
  • Object detection model (using YOLO V2): Find the figure body recognized in the picture, and find its coordinates and size at the same time

System environment

First, a computer with Linux system is required If your main system is Windows, you can use the following system environment:

  • You can use a virtual machine, virtual box or vmware, the system recommends installing Ubuntu20.04
  • Or install dual systems, please search and learn the installation method yourself, or see this dual system installation tutorial

You may want to develop under Windows, but it is strongly recommended to use Linux instead of Windows:

  • First of all, most model training frameworks support Linux first, and the difficulty of developing under Linux will be easier than developing under Windows
  • As a developer, learning to use Linux is a basic skill. Of course, unless you are a fan of Windows, then I believe you must have the ability to port software from other systems to Windows

Software Installation

Training can use CPU for training, but the speed is relatively slow. If you use a dedicated graphics card (GPU) for acceleration, the speed will be much faster. Individuals generally use Nvidia graphics cards, such as RTX 3090, of course, use ordinaryGTX 1060 6G memoryversion can be used happily

For the first contact, it is recommended to use the CPU for training first, the environment installation will be much easier, the following only talks about the method of CPU training, GPU please learn by yourself

For GPU usage, please refer to the official Tensorflow GPU Usage Tutorial. If you encounter problems with the graphics card driver, please refer to here, if you encounter problems with docker installation, you can also see here

The following method of use is excerpted from the warehouse’s README, if there are discrepancies, please refer to the warehouse’s README, pay attention to distinguish

  • Clone the training code to local
git clone https://github.com/sipeed/maix_train --recursive
  • Installation dependencies
cd maix_train
pip3 install -r requirements.txt

Chinese users can use Alibaba Cloud or Tsinghua's source, the download speed is faster

pip3 install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/
  • Download nncase v0.1.0-rc5 and unzip it to maix_train/tools/ncc/ncc_v0.1, guaranteed The path of the execution file is maix_train/tools/ncc/ncc_v0.1/ncc

  • Configuration project

Initialize the project first

python3 train.py init

Then edit the maix_train/instance/config.py configuration according to your hardware situation

Prepare the data set

Prepare the data set, the image size is 224x224, the format can refer to the data set example under maix_train/datasets

See also Maxhub's data set requirements

Train classification model

python3 train.py -t classifier -z datasets/test_classifier_datasets.zip train

Or unzip the data set to a folder, specify the data set folder

python3 train.py -t classifier -d datasets/test_classifier_datasets train

Train the target detection model

python3 train.py -t detector -z datasets/test_detector_xml_format.zip train

Use model

Like the model trained with Maixhub, a zip file will be generated in the out folder, which contains the results, copy all files to the root directory of the SD card, and then power on the development board Just run