You know when you have some important stuff to do, but you try so hard to find excuses and other things to do to postpone everything? Today, 28/01/2020, it's one of those days :)) So, I really wanted to loose some time and explore a little bit these new edge AI devices made by Google that are in my office.
In order to check their performance, I trained a very simple and dumb CNN (feel free to improve it) on the Rock-Paper-Scissor dataset and I made it run on a Coral Dev Board, a Raspberry 4 with the USB Coral Accelerator (connected to USB 3 port) and Raspberry 3 with the USB Coral Accelerator (USB 2 port). Note that the results on the USB Accelerator have been achieved installing the Edge TPU runtime with maximum operating frequency. The inference time is the mean value over 1000 predictions.
These are my results:
| Device | Inference Time [ms] | FPS |
|---|---|---|
| Coral Dev Board | 1.3 | 747 |
| Raspi 4 + USB Accelerator (USB 3) | 1.4 | 710 |
| Raspi 3 + USB Accelerator (USB 2) | 5.0 | 200 |
Side Note fsalv will be or is (depending on when you are reading) a contributor of this repository. He wants to improve it and make it ukulele friendly.
Clone this repository
git clone https://github.com/EscVM/RPS_with_Edge_TPUPython3 is required. I used TensorFlow 2.x for the training, but I uploaded also all converted and original weights. So, if you don't want to re-train the network you can simply use the inference code.
Install on the hosting device the following libraries to make the inference code work :
- opencv-python. N.B. We installed OpenCV4.0 on the Dev Board using this guide as reference.
- numpy
- TensorFlow Lite Interpreter. If you're using the Coral USB Accelerator with the Raspberry download ARM32.
N.B. If you are using the Dev Board, both the Interpreter and the EdgeTPU Compiler are already installed during OS flashing.
Open your terminal in the project folder and launch:
python3 rpc_webcam.pyEnjoy the network predicting the shape of your beautifull hands :)
As I already wrote in the introduction, I made this project very quickly to check the performance of my two Coral devices. So, I didn't spend time building a cool network. If you want to improve the CNN structure or using transfer learning to retrain your prefered architecture, in the project folder you can find the two jupyter notebook I used to train the network and convert from TensorFlow to TFLite. Then you have to use the TPU compiler to make your TFLite file TPU compatible. It's a long, but not difficult process. Here you can find a beutifull summary of the entire chain.