NNoM is a high-level inference Neural Network library specifically for microcontrollers.
Highlights
- Deploy Keras model to NNoM model with one line of code.
- Support complex structures; Inception, ResNet, DenseNet, Octave Convolution...
- User-friendly interfaces.
- High-performance backend selections.
- Onboard pre-compiling - zero interpreter performance loss at runtime.
- Onboard evaluation tools; Runtime analysis, Top-k, Confusion matrix...
The structure of NNoM is shown below:
More detail avaialble in Development Guide
Discussions welcome using issues. Pull request welcome. QQ/TIM group: 763089399.
Recurrent Layers (RNN) (0.4.1)
Recurrent layers (Simple RNN, GRU, LSTM) are implemented in version 0.4.1. Support statful
and return_sequence
options.
New Structured Interface (0.4.0)
NNoM has provided a new layer interface called Structured Interface, all marked with _s
suffix. which aims to use one C-structure to provided all the configuration for a layer. Different from the Layer API which is human friendly, this structured API are more machine friendly.
Per-Channel Quantisation (0.4.0)
The new structred API supports per-channel quantisation (per-axis) and dilations for Convolutional layers.
New Scripts (0.4.0)
From 0.4.0, NNoM will switch to structured interface as default to generate the model header weights.h
. The scripts corresponding to structured interfaces are nnom.py
while the Layer Interface corresponding to nnom_utils.py
.
NNoM is released under Apache License 2.0 since nnom-V0.2.0. License and copyright information can be found within the code.
The aims of NNoM is to provide a light-weight, user-friendly and flexible interface for fast deploying on MCU.
Nowadays, neural networks are wider, deeper, and denser.
[1] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A. (2015). Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1-9).
[2] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).
[3] Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700-4708).
After 2014, the development of Neural Networks are more focus on structure optimising to improve efficiency and performance, which is more important to the small footprint platforms such as MCUs. However, the available NN libs for MCU are too low-level which make it sooooo difficult to use with these complex strucures.
Therefore, we build NNoM to help embedded developers for faster and simpler deploying NN model directly to MCU.
NNoM will manage the strucutre, memory and everything else for the developer. All you need to do is feeding your new measurements and getting the results.
NNoM can be installed as a Python package
pip install git https://github.com/majianjia/nnom@master
NNoM requires Tensorflow version <= 2.14
.
There are multiple options for how to install this, see the TensorFlow documentation.
For example:
pip install 'tensorflow-cpu<=2.14.1'
NOTE: Tensorflow 2.14 supports up until Python 3.11. However, Python 3.12 is not supported.
The C headers and source code in NNoM are distributed in the nnom_core
Python package.
You can find its location by running the following command.
python -c "import nnom_core; print(nnom_core.__path__[0])"
In your build system, add the inc/
and port/
directories as include directories,
and compile the the src/*.c
files.
Guides
RT-Thread-MNIST example (Chinese)
There are many articles compared NNoM with other famous MCU AI tools, such as TensorFlow LiteSTM32Cube.AI .etc.
Raphael Zingg etc from Zurich University of Applied Sciences compare nnom with tflite, cube, and e-Ai in their paper "Artificial Intelligence on Microcontrollers" blog https://blog.zhaw.ch/high-performance/2020/05/14/artificial-intelligence-on-microcontrollers/
Butt Usman Ali from POLITECNICO DI TORINO, did below comparison in the thesis: On the deployment of Artificial Neural Networks (ANN) in low cost embedded systems
Both articles shows that NNoM is not only comparable with other popular NN framework but with faster inference time and sometime less memory footprint.
Note: These graphs and tables are credited to their authors. Please refer the their original papers for details and copyright.
Documented examples
Please check examples and choose one to start with.
*Notes: NNoM now supports both HWC and CHW formats. Some operation might not support both format currently. Please check the tables for the current status. *
Core Layers
Layers | Struct API | Layer API | Comments |
---|---|---|---|
Convolution | conv2d_s() | Conv2D() | Support 1/2D, support dilations (New!) |
ConvTransposed (New!) | conv2d_trans_s() | Conv2DTrans() | Under Dev. |
Depthwise Conv | dwconv2d_s() | DW_Conv2D() | Support 1/2D |
Fully-connected | dense_s() | Dense() | |
Lambda | lambda_s() | Lambda() | single input / single output anonymous operation |
Batch Normalization | N/A | N/A | This layer is merged to the last Conv by the script |
Flatten | flatten_s() | Flatten() | |
Reshape (New!) | reshape_s() | N/A | |
SoftMax | softmax_s() | SoftMax() | Softmax only has layer API |
Activation | N/A | Activation() | A layer instance for activation |
Input/Output | input_s()/output_s() | Input()/Output() | |
Up Sampling | upsample_s() | UpSample() | |
Zero Padding | zeropadding_s() | ZeroPadding() | |
Cropping | cropping_s() | Cropping() |
RNN Layers
Layers | Status | Struct API | Comments |
---|---|---|---|
Recurrent NN Layer(New!) | Alpha | rnn_s() | Layer wrapper of RNN |
Simple Cell (New!) | Alpha | simple_cell_s() | |
GRU Cell (New!) | Alpha | gru_cell_s() | Gated Recurrent Network |
LSTM Cell (New!) | Alpha | lstm_s() | Long Short-Term Memory |
Activations
Activation can be used by itself as layer, or can be attached to the previous layer as "actail" to reduce memory cost.
There is no structred API for activation currently, since activation are not usually used as a layer.
Actrivation | Struct API | Layer API | Activation API | Comments |
---|---|---|---|---|
ReLU | N/A | ReLU() | act_relu() | |
Leaky ReLU (New!) | N/A | LeakyReLU() | act_leaky_relu() | |
Adv ReLU(New!) | N/A | N/A | act_adv_relu() | advance ReLU, Slope, max, threshold |
TanH | N/A | TanH() | act_tanh() | |
Hard TanH (New!) | N/A | TanH() | backend only | |
Sigmoid | N/A | Sigmoid() | act_sigmoid() | |
Hard Sigmoid (New!) | N/A | N/A | N/A | backend only |
Pooling Layers
Pooling | Struct API | Layer API | Comments |
---|---|---|---|
Max Pooling | maxpool_s() | MaxPool() | |
Average Pooling | avgpool_s() | AvgPool() | |
Sum Pooling | sumpool_s() | SumPool() | |
Global Max Pooling | global_maxpool_s() | GlobalMaxPool() | |
Global Average Pooling | global_avgpool_s() | GlobalAvgPool() | |
Global Sum Pooling | global_sumpool_s() | GlobalSumPool() | dynamic output shift |
Matrix Operations Layers
Matrix | Struct API | Layer API | Comments |
---|---|---|---|
Concatenate | concat_s() | Concat() | Concatenate through any axis |
Multiple | mult_s() | Mult() | |
Addition | add_s() | Add() | |
Substraction | sub_s() | Sub() |
NNoM now use the local pure C backend implementation by default. Thus, there is no special dependency needed.
However, You will need to enable libc
for dynamic memory allocation malloc(), free(), and memset()
. Or you can port to the equivalent memory method in your system.
CMSIS-NN/DSP is an optimized backend for ARM-Cortex-M4/7/33/35P. You can select it for up to 5x performance compared to the default C backend. NNoM will use the equivalent method in CMSIS-NN if the condition met.
Please check Porting and optimising Guide for detail.
The script currently does not support implicit act:
x = Dense(32, activation="relu")(x)
Use the explicit activation instead.
x = Dense(32)(x)
x = Relu()(x)
- Attaching an BatchNormalization after each convolutional layer limit the activation range thus help quantisation. BN add no extra computation in NNoM.
- Dont train too much epoch. Large epoch number increases extreme number in activation -> lower the quantisation resolution.
- Leave enough data for bottleneck - do not compress data at before the output of a model, infomation will be lost when it is quantised.
Jianjia Ma [email protected]
Also find me for field supports.
Please contact me using above details if you have any problem.
Example:
@software{jianjia_ma_2020_4158710,
author = {Jianjia Ma},
title = {{A higher-level Neural Network library on Microcontrollers (NNoM)}},
month = oct,
year = 2020,
publisher = {Zenodo},
version = {v0.4.2},
doi = {10.5281/zenodo.4158710},
url = {https://doi.org/10.5281/zenodo.4158710}
}