djzenma / tflite-applications-and-models

TFLite applications: Optimized .tflite models (i.e. lightweight and low latency) and code to run directly on your Microcontroller!

Machine Learning on Embedded Systems

TFLite applications: Optimized .tflite models (i.e. lightweight and low latency), tips and tricks to optimize your model and to make it require only about 6KB of memory, and code to run directly on your Microcontroller!

This repo is composed of the following:

Directory	Description
BenchmarkingTool	Tool to benchmark & analyze any TFLite model in terms of: Latency Accuracy Number of Weights Weights memory size Tensors Used Every Tensor's memory allocation
Models_Analysis	Benchmark of the TFLite models mentionned below in excel sheets
Models/MotionSense	TF Models, with different options of quantization, infering depending on the data of grabbed from the accelerometer whether you are going: upstairs/downstairs walking/jogging standing/sitting
Models/MusicClassifier	TF Models, with different options of quantization, predicting one of the following music genres being played via the microphone: Blues Classical Country Disco Hip Hop Jazz Metal Pop Reggae Rock
Models/KeywordSpotting	TF Models, with different options of quantization, predicting whether you are saying the wake word "on!" to the microphone
Models/CarSensor	TF Models, with different options of quantization, predicting whether you are riding a vehicle or not, depending on the data of the accelerometer

---

Model	Weights Size (KB)	Accuracy (%)
MotionSense	6.02	92.31
MusicClassifier	42	88.45
KeywordSpotting	6.01	98.21
CarSensor	2.57	95.00

---

Model Reduction (Optimization) Methodology

Building Blocks

All of the applications mentionned above are time series data, and therefore CNN architectures were suitable for all of them. Most CNNs have kind of a similar structure as this one below:
Input Layer -> (Conv -> MaxPoolingxPi)xN -> (Dense Layer -> Regularizor)xM -> Output Layer

where N = Number of Conv Blocks, M = Number of Dense Blocks, Pi = Number of Pooling Layers in the ith Conv Block.

An important trick in reducing the number of weights in a CNN is adding Pooling Layers as it outputs half the number of its input weights.

Now that we have our building blocks, we can define our optimization algorithm.

Algorithm

Optimizing the CNN network can be done by reducing N and M, and varying Pi of every ith Conv Block until we are satisfied with a good compromise of accuracy and weights size values. In fact, applying a search algorithm on these 2 hyperparameters while having as score the accuracy of the network and the weights size leads to the reduced optimized model. The final part is to find a good loss function to feed to the search algorithm for it to minimize it. For now a weighted average of both accuracy and weights size will do the job, but it might be subject to improvement.

Unsupported Operations

Your original model (before converting to TFLite) might be using operations (ops) that are not supported yet by TFLite, i.e. not the built in TFLite ops. These unsupported ops fall under 1 of 2 categories: Select Ops or Custom Ops. If you are using one of them, then your converted TFLite model will not be lightweight because it didn't optimize these operations.

Select Ops are ops that are supported by TF but not yet by TFLite (keep in mind that it is still in the experimental phase) and Custom Ops are ops that you implemented yourself (and not built in in TF). To find out if your model uses one of the select ops, you can either visit the documentation of TFLite where they have a list of all of their built in ops, or you can find out during the conversion process to TFLite. To clarify this latter option, when you are converting your model to TFLite, if the model converts successfully without the following line:

converter.target_spec.supported_ops = [tf.lite.OpsSet.SELECT_TF_OPS]

then your model does not use any Select Ops. If an error occurs, try putting this line, if it converts then you are using a Select Op. If it doesn't convert both ways, then you are most probably using a Custom Op.

To know how to convert a model with Select Ops and for further details on Select Ops, refer to TF's documentation.

To know how to convert a model with Custom Ops and for further details on Custom Ops, refer to TF's documentation.

Example of a SELECT operation

This example was found empirically and I don't think that it is mentionned anywhere in the TF documentation. It is the data_format attribute found in the tf.keras.layers.Conv2D component. If you are using it then remove it (the default is the channels_last option) and don't overwrite the default to channels_first. This way, your model will be optimized.