Exploring Mobile-Friendly Machine Learning Frameworks: From ncnn to PaddlePaddle

Just Chatting

Hello everyone, I’m the person who has been working from home for half a month and will continue for another week. Working from home feels dull and makes me miss the office; we often want the opposite situation. Our generation has been PUA‑ed for so long that we can’t stay idle. Recently many neighborhoods in Beijing are demanding a lift of restrictions, and I wonder if I’m the only one worrying about elders and children at home. Today I want to share some thoughts and the technologies I’ve been looking at.

From a technical perspective, business constantly chases technology, and technology in turn empowers business. Because I work in mobile and front‑end technology management, I keep an eye on emerging mobile technologies and consider which might become core competitive advantages. I recalled a jump‑rope app that used a machine‑learning model with a .nb file, which I never examined closely. My company also has a photo‑based homework‑grading product that is costly and unmaintained. I decided to explore machine‑learning basics to see if “edge intelligence” could reduce costs and broaden my skill set.

What Is Machine Learning?

Machine learning is an interdisciplinary field involving probability, statistics, approximation theory, convex analysis, and algorithmic complexity. It studies how computers can simulate or implement human learning to acquire new knowledge or skills, reorganize existing knowledge, and continuously improve performance.

Common Machine Learning Frameworks

ncnn (used)

ncnn is a high‑performance neural‑network inference framework optimized for mobile devices. I used it when developing a handwritten‑recognition feature, converting a Caffe‑trained model to ncnn format, which worked exceptionally well on mobile.

Caffe (used)

Caffe (Convolutional Architecture for Fast Feature Embedding) is a deep‑learning framework known for expressiveness, speed, and modularity, developed by Berkeley AI Research. I trained the MNIST dataset with it and achieved about 98% accuracy.

TensorFlow

Google’s TensorFlow has a rich ecosystem, suitable for beginners and also offers server‑side implementations. Keras, the official Python front‑end for TensorFlow, provides a simple API for rapid machine‑learning development.

PyTorch

Developed by Facebook, PyTorch is an open‑source Python library for machine learning, widely used in natural‑language processing and computer‑vision projects. I also explored YOLOv5, which gained popularity after a global wheat‑detection Kaggle competition.

PaddlePaddle

PaddlePaddle, Baidu’s deep‑learning platform (known as “飞桨”), powers the jump‑rope app mentioned earlier. Its main drawback is difficulty converting models to other frameworks or to mobile‑ready formats.

Conclusion

That’s all for now. If the article attracts enough interest, I’ll write a follow‑up on how to convert models between frameworks or how to implement an automatic OCR system for grading arithmetic homework.

Are you interested in machine‑learning technologies? Follow my public account (闲坐说) and leave a comment. Thanks for reading.