Multi-Component Optimization of Neural-Networks for IoT Devices
Notebook for an end-to-end multi-component NN optimization sequence to enable the execution of high memory plus computation demanding models on MCUs, small CPUs, and AIOT boards. Post multi-component optimization, the resultant models are smaller in size, consume less power when execution, and show low latency. Our sequence is generic and applicable to any state-of-the-art models trained for anomaly detection, predictive maintenance, machine vision, etc.
Note: The .ipynb files can be loaded and viewed from the Github page, but it needs to be reloaded a couple of times as the file is big. Hence, it is best to download and open via Google Colab or Jupyter Notebook, thanks.
Experiment: Multi-component Optimization of CNNs
In the notebooks, we use the standard MNIST Fashion (produces CNN1) and MNIST Digits (produces CNN2) datasets to train a basic CNN whose architecture is shown below
Both these datasets are imported via the tf.keras.dataset.name function with its default train and test sets. After importing, we apply all suitable optimizers before, during, and after training CNNs and analyze the memory conservation, accuracy, and inference speedups. In the following, we brief each optimization component whose implementations are provided in the notebooks.
Pre-training Optimization
We first apply the pruning technique on CNNs and present the changes in inference time and size in below Figure c. Similarly, we also perform quantization-aware training of CNNs and show the changes in below Figure b.
Post-training Optimization
We performed Int with float fallback quantization on original CNNs and show its architecture and performance in below Figure d.
Here we quantized the original CNN's Float32 weights and activations to Float16 values. Users can use this Float16 quantization when they want to achieve reasonable compression rates (we obtain approx. 6x compression) without loss of precision (we experience only 0.01 % loss in accuracy). Also, Float16 models run on small CPUs without modification. In below Figure e, we show the Float16 quantized model's architecture, inference time, and size changes.
We also converted the CNN's weights & activation to 8-bit integers and show its architecture, inference time, and size changes below Figure f. Here, the size reduced and inference time improved since, after quantization, the inference is carried out using Integer-only arithmetic.
Operations and Graph Optimization
When designing ML models for tiny IoT hardware, only limited operations can be used to keep the cost low. Over 90% of arithmetic operations are used by convolutional (CONV) layers. So, we already convert floating-point operations into int-8 (fixed point) during post-training quantization. Here, we decompose (depthwise separation) the 2-D CONVs, followed by 1-D CONVs, aiming to reduce parameters and operations count. When using this depth-separation concept on 3D filters, a regular 3D convolution uses C * A * B multiplications, whereas a depth-separable 3D convolution only requires C + A + B multiplications.
In below Figure g, we show the operations optimized model's architecture, inference time, and size changes.
We implemented and performed all applicable arithmetic simplification rewrites and graph structure optimization tasks on CNNs and in the above Figure h, we show the graph optimized model's inference time and size changes.
Results Analysis
We performed analysis based on the experiment results and report the best optimization sequence for:
Smallest Model Size: When users want the smallest possible trained model, we recommend the Graph optimized then integer with float fallback quantized version. This model is only 22.5 KB, i.e., 12.06 x times smaller than the original CNN.
Accuracy Preservation: When the target hardware can accommodate a few extra KB, naturally we would try to fit the top-performing model. In such cases, we recommend to load and use the Graph optimized then integer only quantized version. For MNIST Fashion, the accuracy increased by 0.27 % and by 0.13 % for MNIST Digits.
Fast Inference: For real-time applications, we naturally tend to load and use the fastest inference results-producing models. In such cases, we recommend the Operations optimized then float16 quantized version that produces the fastest unit inference results in 0.06 ms.
If the code is useful, please consider citing Train++ and ML-MCU papers using the BibTex entry below.
@inproceedings{sudharsan2022multicomponent,
title = {Multi-Component Optimization and Efficient Deployment of Neural-Networks on Resource-Constrained IoT Hardware},
author = {Sudharsan, Bharath and Sundaram, Dineshkumar and Patel, Pankesh and Breslin, John G. and Ali, Muhammad Intizar and Dustdar, Schahram and Zomaya, Albert and Ranjan, Rajiv},
url = {https://arxiv.org/abs/2204.10183},
booktitle = {arXiv},
year = {2022}
}