alessandrocapotondi/mobilenet_v1_stmcube_ai

Mobilenet v1 (3,128,128, alpha=0.25) on STMH7 using STMCube AI

Mobilenet on STMH7 using STMCubeMX.AI

(new version is located here: https://github.com/alessandrocapotondi/MobileNet_v1_x_cube_ai_4.1.0)

The repo contains a STMWorkbench project that aims to fit a Mobilenet v1 (3,128,128, alpha=0.25) into a MCU STM32H7 using the STMCubeMX.AI flow. Note that the selected Mobilenet version is the biggest model that can be fitted on the MCU STM32H7 using the STMCubeMX.AI flow.

This project will be used as reference benchmark of our Quantized Mobilenet based onCMSIS-NN: https://github.com/EEESlab/mobilenet_v1_stm32_cmsis_nn

Current Status

Mobilenet v1 (3,128,128, alpha=0.25) compiles and works on STMicroelectronics STM32 Nucleo-144.

Content

The project folder structure follows the classic STMWorkbench project template. In addition it contains the code generated by STMCubeMX.AI for the Mobilenet included inside the folder keras_model/:

• Drivers/: (generated) The Folder contains the ARM CMSIS and additional drivers for the MCU;
• Middlewares/ST/AI/AI/: (generated) STMCubeMX.AI generated code for the Mobilenet.
• Src/: (generated) Sources.
• Inc/usr_mobilenet.h: user-space header file for Mobilenet application execution.
• Src/usr_mobilenet.c: user-space applications and utilities for Mobilenet application execution.
• keras_model/: The folder contains a Python notebook used to fix and convert the Keras Mobilenet pre-trained model to a model that can be imported into the STMCubeMX.AI flow. According to the discussion [1], keras.advanced_activations.ReLU layers are not usable directly in STCubemx.ai. The function remove such layers and it substitutes to supported keras.activations.relu.
• mobilenet_cubemx_ai.ioc: STMCubeMX project file.

Measured Performances

Million MACs	Million Parameters	Top-1 Accuracy	Top-5 Accuracy	CPU Cycles (MCycles)	Latency @400MHz (s)	MMACs/s	MMACs/s/W*
14	0.47	41.5	66.3	99	0.247	56.7	84.6

• This number refers to an average power consumption of 0.68 Watt measured at the power source, so it CANNOT be considered as efficiency metric of the MCU, but it is the efficiency of this particular solution, which includes LED activations, GPIO triggering, and all the other peripherals on the board.

Execution Screenshot

mobilenet_v1_stmcube_ai's Issues

Is benchmark available?

I noticed you were going to benchmark CubeAI solution you used in this project with CMSIS NN solution you used in your other project. Do you have the first results? I'm also interested in CubeAI vs CMSIS NN.

Currently, CubeAI supports float32 only. But they are planning to support quantization as well in the near future.

MobileNetV2

Thanks for the example, but a number of questions arise:

1. Will such a method work correctly on an already trained MobileNet model ? After all, we violated the internal architecture of the model. Or do I need to train this model again ?

2. You have an example on the MobileNet model. Will your method work for MobileNetV2 ?