This time, I was fortunate to apply for the book "Deep Learning" and read it for three weeks. Since I used to use single-chip microcomputers for development, I hope to acquire some relevant knowledge about machine learning on single-chip microcomputers. The book's description of deep learning content is vivid and uses many examples close to life, allowing people to intuitively feel the principles of deep learning. Through reading and consulting materials, I hope to share with you some of the gains I have made on the issues I am concerned about. The first is the solution to the acquisition of data sets in machine learning. Usually we need a certain amount of data to complete the training, and before that we usually need to manually generate data sets to meet usage needs. Through learning and consulting, I also learned the following acquisition methods:
Public datasets : Many research institutions and universities release some standard public datasets, which are usually cleaned and annotated and suitable for academic research and educational purposes. For example, the MNIST dataset is an entry-level dataset for handwritten digit recognition, CIFAR-10 and CIFAR-100 are commonly used image recognition datasets, ImageNet is a large-scale image recognition dataset, and Pascal VOC is often used for object detection and image segmentation tasks.
Web crawler : You can crawl data from the Internet by writing a web crawler program. This method can obtain a large amount of data, but it is necessary to deal with the uncertainty and quality of the data.
Data platforms : Some data platforms such as Kaggle, Google Dataset Search, AWS Public Datasets, etc. provide a large number of data sets covering various fields and applications.
When developing a project, there are some cases where you need to deploy a machine learning model on the microcontroller, and STM32Cube.AI is usually used for model conversion and code generation. When I use a microcontroller for machine learning, I also understand some things that need to be paid attention to to optimize performance, such as the following ways: Use bit operations instead of arithmetic operations : Bit operations are usually faster than arithmetic operations, such as using bitwise AND, bitwise OR, bitwise XOR and other operations to process flag bits, or using shift operations instead of multiplication and division. Avoid unnecessary operations : Remove unnecessary variable assignments and function calls in the code to reduce program complexity and execution time. Control data type size : Try to use smaller data types, such as charor int, instead of float, because floating-point operations usually require more CPU cycles and memory space. Utilize hardware resources : For example, use DMA (direct memory access) to transfer data to reduce CPU load, or use the microcontroller's hardware accelerator (if it exists) to perform specific tasks. Self-increment and self-decrement instructions : Use self-increment and self-decrement operations such i++as i--and, because some compilers optimize these operations and are more efficient. Optimize algorithms : Select or design algorithms that are suitable for the resource limitations of the microcontroller, such as using simple linear regression instead of complex neural network models. Code reuse : Reuse code through functions or macros to reduce code redundancy and improve code maintainability. Register optimization : Store frequently used variables in registers to speed up access. Instruction pipeline optimization : By organizing the code reasonably, instructions can be executed in parallel to improve execution efficiency. Memory optimization : Optimize data structures and memory usage, reduce memory fragmentation, and increase memory access speed. Interrupt optimization : Reasonably configure interrupts, reduce the complexity of interrupt handlers, and ensure that interrupt service routines are as short as possible. Model optimization : During the model training stage, the model size can be reduced through model compression, quantization and other technologies to make it more suitable for running on a microcontroller.
I wonder if any of my colleagues have rich experience in using microcontrollers to deploy machine learning models. I hope we can communicate and learn from each other, as I am still relatively unfamiliar with this area.
提升卡
变色卡
千斤顶
京公网安备 11010802033920号
