README.md

---
library_name: keras
tags:
- timeseries
---

# Timeseries classification from scratch

Based on the _Timeseries classification from scratch_ example on [keras.io](https://keras.io/examples/timeseries/timeseries_classification_from_scratch/) created by [hfawaz](https://github.com/hfawaz/).

## Model description

The model is a Fully Convolutional Neural Network originally proposed in [this paper](https://arxiv.org/abs/1611.06455).
The implementation is based on the TF 2 version provided [here](https://github.com/hfawaz/dl-4-tsc/).
The hyperparameters (kernel_size, filters, the usage of BatchNorm) were found via random search using [KerasTuner](https://github.com/keras-team/keras-tuner).

## Intended uses & limitations

Given a time series of 500 samples, the goal is to automatically detect the presence of a specific issue with the engine.

The data used to train the model was already _z-normalized_: each timeseries sample has a mean equal to zero and a standard deviation equal to one.

## Training and evaluation data

The dataset used here is called [FordA](http://www.j-wichard.de/publications/FordPaper.pdf). The data comes from the [UCR archive](https://www.cs.ucr.edu/%7Eeamonn/time_series_data_2018/). The dataset contains:

 - 3601 training instances
 - 1320 testing instances
 
 Each timeseries corresponds to a measurement of engine noise captured by a motor sensor.
 
## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:

| name | learning_rate | decay | beta_1 | beta_2 | epsilon | amsgrad | training_precision |
|----|-------------|-----|------|------|-------|-------|------------------|
|Adam|9.999999747378752e-05|0.0|0.8999999761581421|0.9990000128746033|1e-07|False|float32|

 ## Model Plot

<details>
<summary>View Model Plot</summary>

![Model Image](./model.png)

</details>


<center>
Model reproduced by <a href="https://github.com/EdAbati" target="_blank">Edoardo Abati</a>
</center>