NNoM is a high-level linference Neural Network library specifically for microcontrollers.
[English Manual] [Chinese Intro]
Highlights
- Deploy Keras model to NNoM model with one line of code.
- User-friendly interfaces.
- Support complex structures; Inception, ResNet, DenseNet, Octave Convolution...
- High-performance backend selections.
- Onboard (MCU) 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.
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.
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 is now working closely with Keras (You can easily learn Keras in 30 seconds!). There is no need to learn TensorFlow/Lite or other libs.
Guides
RT-Thread-MNIST example (Chinese)
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 | HWC | CHW | Layer API | Comments |
|---|---|---|---|---|
| Convolution | ✓ | ✓ | Conv2D() | Support 1/2D |
| Depthwise Conv | ✓ | ✓ | DW_Conv2D() | Support 1/2D |
| Fully-connected | ✓ | ✓ | Dense() | |
| Lambda | ✓ | ✓ | Lambda() | single input / single output anonymous operation |
| Batch Normalization | ✓ | ✓ | N/A | This layer is merged to the last Conv by the script |
| Flatten | ✓ | ✓ | Flatten() | |
| SoftMax | ✓ | ✓ | SoftMax() | Softmax only has layer API |
| Activation | ✓ | ✓ | Activation() | A layer instance for activation |
| Input/Output | ✓ | ✓ | Input()/Output() | |
| Up Sampling | ✓ | ✓ | UpSample() | |
| Zero Padding | ✓ | ✓ | ZeroPadding() | |
| Cropping | ✓ | ✓ | Cropping() |
RNN Layers
| Layers | Status | Layer API | Comments |
|---|---|---|---|
| Recurrent NN | Under Dev. | RNN() | Under Developpment |
| Simple RNN | Under Dev. | SimpleCell() | Under Developpment |
| Gated Recurrent Network (GRU) | Under Dev. | GRUCell() | Under Developpment |
Activations
Activation can be used by itself as layer, or can be attached to the previous layer as "actail" to reduce memory cost.
| Actrivation | HWC | CHW | Layer API | Activation API | Comments |
|---|---|---|---|---|---|
| ReLU | ✓ | ✓ | ReLU() | act_relu() | |
| TanH | ✓ | ✓ | TanH() | act_tanh() | |
| Sigmoid | ✓ | ✓ | Sigmoid() | act_sigmoid() |
Pooling Layers
| Pooling | HWC | CHW | Layer API | Comments |
|---|---|---|---|---|
| Max Pooling | ✓ | ✓ | MaxPool() | |
| Average Pooling | ✓ | ✓ | AvgPool() | |
| Sum Pooling | ✓ | ✓ | SumPool() | |
| Global Max Pooling | ✓ | ✓ | GlobalMaxPool() | |
| Global Average Pooling | ✓ | ✓ | GlobalAvgPool() | |
| Global Sum Pooling | ✓ | ✓ | GlobalSumPool() | A better alternative to Global average pooling in MCU before Softmax |
Matrix Operations Layers
| Matrix | HWC | CHW | Layer API | Comments |
|---|---|---|---|---|
| Concatenate | ✓ | ✓ | Concat() | Concatenate through any axis |
| Multiple | ✓ | ✓ | Mult() | |
| Addition | ✓ | ✓ | Add() | |
| Substraction | ✓ | ✓ | Sub() |
NNoM now use the local pure C backend implementation by default. Thus, there is no special dependency needed.
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:
Dense(32, activation="relu")
Use the explicit activation instead.
Dense(32)
Relu()
Jianjia Ma
Please contact us using above details.