Accelerate Deep Learning on Raspberry Pi 4 CPU
We are a group of volunteers aiming to start up with ** embedded SW optimization technology ** as the core competence to bring out the HW performance of ** multi-core CPU ** and ** SIMD architecture **.
I am challenging how much Deep Learning can be accelerated with only ** CPU ** of Raspberry Pi 3/4.
In the past, I was targeting frameworks such as Chainer and darknet, but now I am trying to speed up the ONNX runtime.
The results at this time are as follows.
@onnxruntime on RPi4(CPU Only)
— Project-RAIZIN (@ProjectRaizin) September 8, 2020
MobileNetV3(Image clasification)
MobileNetV2-SSDLite(Image detection)
Original vs. Accelerated#RaspberryPi #Python #DeepLearninghttps://t.co/wvBLn9Tfes
Originally, Microsoft and Facebook are promoting the project, so it is difficult to speed up several times, but I managed to double the performance by tuning im2col, gem, Activation function, etc.
In addition, we have released demo videos of various models. Youtube channel
Our commitment points
- 1. Perform Deep Learning using only the CPU of the Raspberry Pi
- 2. Using 32-bit version Rasbian
- Not yet compatible with 64-bit Raspbian ...
- 3. Use the existing model as it is
- We do not reduce the weight such as quantization / pruning / distillation (cannot be done in the first place ...)
- 4. Do not change the calculation accuracy as much as possible
Acceleration approach
The acceleration approach is common as shown below.
- a. Code optimization
- b. Multi-core parallelization
- c. SIMD vectorization
- d. Software Pipelining
- e. Memory efficiency
- b. Multi-core parallelization
I think that it is a characteristic of us that there is no other attitude to squeeze a general item like ** a little faster ** a little faster ** while taking a profile.
At the end of the article
This time, I have only introduced the results, but I would like to summarize the technical data for each item as a memorandum and publish it as needed.
Recommended Posts