This repository contains the slides presented and code used of the tutorial presented at ECAI 2024:
"Tutorial on Improving Deep Learning by Exploiting Synthetic Images"
Work in groups to try and surpass the baseline set by classical models on the Boston housing dataset.
Lazypredict - refer to this notebook: Notebooks/Lazypredict/LazyPredict_Regression.ipynb
Using synthetic images, experiment with either vision models like CNNs or ViTs, and explore hybrid models. Below are the architectures that will be presented, and the ones you will modify and use during the session:
-
Synthetic images using CNN
-
Synthetic images using Hybrid Neural Network with ViT (HyNNViT)
Here are the notebooks you can directly open and run in Google Colab:
Note: Before running the notebooks, you will need to download the required dataset. For the practical session, we will use a small dataset, specifically the Boston housing dataset, which is located in Data/Regression/boston.csv.
The notebooks listed below are designed for regression tasks:
-
TensorFlow - CNN:
-
TensorFlow - CNN + MLP Hybrid:
-
TensorFlow - Vision Transformer (ViT):
-
TensorFlow - ViT + MLP Hybrid:
In this tutorial, we will explore various methods to transform tabular data into images to take advantage of deep learning models such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs).
TINTOlib is a state-of-the-art library that wraps the most important techniques for the construction of Synthetic Images from Sorted Data (also known as Tabular Data).
Citing TINTO: If you used TINTO in your work, please cite the SoftwareX:
@article{softwarex_TINTO,
title = {TINTO: Converting Tidy Data into Image for Classification with 2-Dimensional Convolutional Neural Networks},
journal = {SoftwareX},
author = {Manuel Castillo-Cara and Reewos Talla-Chumpitaz and Raúl García-Castro and Luis Orozco-Barbosa},
volume={22},
pages={101391},
year = {2023},
issn = {2352-7110},
doi = {https://doi.org/10.1016/j.softx.2023.101391}
}And use-case developed in INFFUS Paper
@article{inffus_TINTO,
title = {A novel deep learning approach using blurring image techniques for Bluetooth-based indoor localisation},
journal = {Information Fusion},
author = {Reewos Talla-Chumpitaz and Manuel Castillo-Cara and Luis Orozco-Barbosa and Raúl García-Castro},
volume = {91},
pages = {173-186},
year = {2023},
issn = {1566-2535},
doi = {https://doi.org/10.1016/j.inffus.2022.10.011}
}All the methods presented can be called using the TINTOlib library. The methods presented include:
| Model | Class | Features | Hyperparameters |
|---|---|---|---|
| TINTO | TINTO() |
blur |
problem algorithm pixels submatrix blur amplification distance steps option random_seed times verbose |
| IGTD | IGTD() |
problem scale fea_dist_method image_dist_method max_step val_step error switch_t min_gain zoom random_seed verbose |
|
| REFINED | REFINED() |
problem n_processors hcIterations zoom random_seed verbose |
|
| BarGraph | BarGraph() |
problem pixel_width gap zoom verbose |
|
| DistanceMatrix | DistanceMatrix() |
problem zoom verbose |
|
| Combination | Combination() |
problem zoom verbose |
|
| SuperTML | SuperTML() |
problem columns font_size image_size verbose |
|
| FeatureWrap | FeatureWrap() |
problem size bins zoom verbose |
|
| BIE | BIE() |
problem precision zoom verbose |
- For more detailed information, refer to the TINTOlib ReadTheDocs.
- GitHub repository: TINTOlib Documentation.
- PyPI: PyPI.