TensorFlores: An Enhanced Python-based TinyML Framework

The TensorFlores framework is a Python-based solution designed for optimizing machine learning deployment in resource-constrained environments. It introduces an evolving clustering-based quantization, enabling quantization-aware training (QAT) and post-training quantization (PTQ) while preserving model accuracy. TensorFlores seamlessly converts TensorFlow models into optimized formats and generates platform-agnostic C++ code for embedded systems. Its modular architecture minimizes memory usage and computational overhead, ensuring efficient real-time inference. By integrating clustering-based quantization and automated code generation, TensorFlores enhances the feasibility of TinyML applications, particularly in low-power and edge AI scenarios. This framework provides a robust and scalable solution for deploying machine learning models in embedded and IoT systems.

• Software description
• Installation
• Usage Examples
• References
• License

---

Dependencies

Python v3.9.6

pip install -r requirements.txt

---

Software description

The TensorFlores framework is a Python-based solution designed for optimizing machine learning deployment in resource-constrained environments

Software architecture

The architecture of TensorFlores can be divided into four primary layers:

• Model Training: A high-level API for the streamlined creation and training of MLP, supporting evolutionary vector quantization during training;

• Json Handle: Responsible for interpreting TensorFlow models and generating structured JSON files, serving as an intermediary representation for both TensorFlow and TensorFlores models;

• Quantization: Dedicated to processing the structured JSON model representation and applying PTQ techniques;

• Code Generation: Responsible to processing the structured representation of the JSON model and generating the machine learning model in C++ format to be embedded in the microcontroller, whether quantised or not.

Software structure

The project directory is divided into key components, as illustrated in Figure:

tensorflores/
├── models/
│   └── multilayer_perceptron.py
├── utils/
│   ├── autocloud/
│   │   ├── auto_cloud_bias.py
│   │   ├── auto_cloud_weight.py
│   │   ├── data_cloud_bias.py
│   │   ├── data_cloud_weight.py
│   │   └── __init__.py
│   ├── array_manipulation.py
│   ├── clustering.py
│   ├── cpp_generation.py
│   ├── json_handle.py
│   ├── quantization.py
│   └── __init__.py

Software functionalities

The pipeline illustrated in Figure outlines a workflow for optimizing and deploying machine learning models, specifically designed for resource-constrained environments such as microcontrollers. The software structure is divided into four main blocks: model training (with or without quantization-aware training), post-training quantization, TensorFlow model conversion, and code generation, which translates the optimized model into platform-agnostic C++ code.

 

 

The parameters are highly customizable, as shown in Table 1, which lists the class parameters and their corresponding default input values

Class Parameters	Type	Input Values
input_size	int	5
hidden_layer_sizes	list	[64, 32]
output_size	int	1
activation_functions	list	'sigmoid', 'relu', 'leaky_relu', 'tanh', 'elu', 'softmax', 'softplus', 'swish', 'linear'
weight_bias_init	str	'RandomNormal', 'RandomUniform', 'GlorotUniform', 'HeNormal'
training_with_quantization	bool	True or False

Table 1 - MLP Initialization Parameters.

The "train" method has the following main parameters:

Parameter	Type	Input Values
X	list	List of input data for training
y	list	List of corresponding labels
epochs	int	Default: 100
learning_rate	float	Default: 0.001
loss_function	str	'mean_squared_error', 'cross_entropy', 'mean_absolute_error', 'binary_cross_entropy'
optimizer	str	'sgd', 'adam', 'adamax'
batch_size	int	Default: 36
beta1	float	Default: 0.9 (Adam first moment)
beta2	float	Default: 0.999 (Adam second moment)
epsilon	float	Default: 1e-7 (Avoid division by zero in Adam)
epochs_quantization	int	Default: 50
distance_metric	str	'euclidean', 'manhattan', 'minkowski', 'chebyshev', 'cosine', 'hamming', 'bray_curtis', 'jaccard', 'wasserstein', 'dtw' and 'mahalanobis'
bias_clustering_method		Clustering method for biases
weight_clustering_method		Clustering method for weights
validation_split	float	Default: 0.2 (Validation data percentage)

Table 2 - Configurable Train Method Parameters.

Table 3 presents a summary of the clustering algorithms and their respective configuration parameters.

Algorithm	Parameter	Value
AutoCloud	Threshold ($m$)	1.414
MeanShift	Bandwidth ($b$)	0.005
	Maximum iterations	300
	Bin seeding	True
Affinity Propagation	Damping ($d$)	0.7
	Maximum iterations	500
	Convergence iterations	20
DBStream	Clustering threshold ($\tau$)	0.1
	Fading factor ($\lambda$)	0.05
	Cleanup interval	4
	Intersection factor	0.5
	Minimum weight	1

Table 3- Clustering Algorithms and Their Respective Parameters.

Installation

You can download our package from the PyPi repository using the following command:

pip install  tensorflores

If you want to install it locally you download the Wheel distribution from Build Distribution.

First navigate to the folder where you downloaded the file and run the following command:

pip install tensorflores-0.1.4-py3-none-any.whl

---

Usage Example

The following four examples will be considered:

Example 01

Implementation and Training of a Neural Network Using TensorFlores:

Example 02

Implementation and Training of a Neural Network with quantization-aware training (QAT) Using TensorFlores:

Example 03

Post-Training Quantization with TensorFlores:

Example 04

Converting a TensorFlow Model using TensorFlores:

Auxiliary

This section provides an example of code that transforms an input matrix (X_test) and (y_test) into a C++ array format.

The Arduino code to deployment are avaliable here:

Other Models

Please check the informations for more information about the other models been implemented in this package.

Literature reference

1. T. K. S. Flores, M. Medeiros, M. Silva, D. G. Costa, I. Silva, Enhanced Vector Quantization for Embedded Machine Learning: A Post-Training Approach With Incremental Clustering, IEEE Access 13 (2025) 17440 17456. doi:10.1109/ACCESS.2025.3532849.

2. T. K. S. Flores, I. Silva, M. B. Azevedo, T. d. A. de Medeiros, M. d. A. Medeiros, D. G. Costa, P. Ferrari, E. Sisinni, Advancing TinyMLOps: Robust model updates in the internet of intelligent vehicles, IEEE Micro (2024) doi:10.1109/MM.2024.3354323.

License

This package is licensed under the MIT License - © 2023 Conect2ai.