Custom Model Generation with MindSpore Lite | HMS ML Kit

Published in

Huawei Developers

6 min readNov 25, 2020

MindSpore is an open-sourced framework for AI based application development which is announced by Huawei. It is a robust alternative to AI frameworks such as TensorFlow and PyTorch which are widely used in the market.

Let’s start by emphasizing the features and advantages of MindSpore framework:

MindSpore implements AI algorithms for easier model development and provides cutting-edge technologies with Huawei AI processors to improve runtime efficiency and computing performance.
One of its advantages is that it can be used in several environments like on devices, cloud and edge environments. It supports operating systems like IOS and Android, and AI applications on various devices such as mobile phones, tablets and IoT devices.
MindSpore supports parallel training across hardware to reduce training times. It maximizes both hardware computing power and minimizes inference latency and power consumption.
It provides dynamic debugging ability for developers which enables to find out bugs in the apps easily.
According to Huawei, MindSpore does not process data by itself but ingests the gradient and model information that has been processed. This ensures the integrity of sensitive data.

MindSpore Lite is an inference framework for custom models which is provided by HMS ML Kit to simplify the integration and development. The developers can define their own model and implement model inference thanks to MindSpore Lite capabilities.

MindSpore Lite is compatible with commonly used AI platforms like TensorFlow Lite, Caffe, and Onnx. Different models can be converted into .ms (MindSpore) format and then run perfectly.
Custom models can be deployed and executed easily since they are compressed and occupy small storage.
It provides complete APIs to integrate inference framework of an on-device custom model.

HMS ML Kit enables to train and generate custom models with deep machine learning. It also offers pre-trained image classification model. You can develop your own custom model by using Transfer Learning feature of ML Kit with a specific dataset.

I will basically explain you how to train your own model over an example which contains three plant categories. We will use a small data set for reference and train the image classification model to identify cactus, pine and succulent plants. The model will be created by using HMS Toolkit plug-in and AI Create.

HMS Toolkit: As a lightweight IDE tool plugin, HMS Toolkit implements app creation, coding, conversion, debugging, test, and release. It helps you integrate HMS Core APIs with lower costs and higher efficiency.

AI Create: Provides the transfer learning capabilities of image classification and text classification. Images and texts can be identified thanks to AI Create. It uses MindSpore as a training framework and MindSpore Lite as inference framework.

Note: Use the Android Studio marketplace to install HMS Toolkit plug-in. Please go to File > Settings > Plugins > Marketplace, enter HMS Toolkit into the search box and click install. After installation complete, restart Android Studio.

We should prepare the environment to train our model first. AI Create only supports Windows operating system currently. Please open Coding Assistant by using the new HMS section that came with HMS Toolkit plug-in. Go to AI > AI Create in Coding Assistant and select Image and click Confirm for Image Classification.

After this step HMS Toolkit automatically downloads resources for you. If the Python environment is not configured, the dialog box will be displayed as below.

Note: You should download and install Python 3.7.5 from the link to use AI Create. After installation complete, please do not forget to add Python installation path into the Path variable in Environment Variables and restart the Android Studio.

After environment is ready, if you select Image and click Confirm from AI Create it will automatically start to install MindSpore. Please be sure the framework has been installed successfully by checking the Event logs.

From now, new model section will be opened to select an image folder to train our own model. You should prepare your data set in accordance with the requirements. We will train the model for our demo to identify cactus, succulent and pine plants with a small data set.

The folder structure should be like below :

The following conditions should be met for image resources:

The minimum number of pictures for each category of training data is 10.
The lower limit of the classification number of the training data set is 2, and the upper limit is 1000.
Supported image formats: .bmp, .jpg, .jpeg, .png or .gif.

After training image folder is selected, please set Output model file path and training parameters. If you check HMS Custom Model, a complete model will be generated. The train parameters affects the accuracy of image recognition model. You can modify them if you have a experience with deep learning. When you click on Create Model, MindSpore will start to train your model according to the data set.

Training process will take a time depending on your data set. As we used a small data set with the minimum number of the requirements it is completed quickly. You can also track training logs, your model will be created on the specified path at the end of the process.

The training results will be shared after model training is completed. AI create enables to test your model by adding the test images before using it in any project. You can also generate a demo project that implements your new model with Generate Demo option.

You should create a new test image folder with the same structure of provided data set.

As you see above, our test average accuracy is calculated as 57.1%. This accuracy can be improved by providing comprehensive data set and training.

You can also use and experience results of your new model over a demo project which can be created by HMS Toolkit. After the demo is created, you may directly run and build the project and check the results on real device.