https://github.com/fmolivato/cwtlayerkeras

tensorflow 2.4 keras layer that compute cwt(continuous wavelet transformation) on signals
https://github.com/fmolivato/cwtlayerkeras

keras tensorflow2 wavelet-transform

Last synced: 23 days ago
JSON representation

tensorflow 2.4 keras layer that compute cwt(continuous wavelet transformation) on signals

• Host: GitHub
• URL: https://github.com/fmolivato/cwtlayerkeras
• Owner: fmolivato
• License: agpl-3.0
• Created: 2021-04-27T07:36:30.000Z (over 3 years ago)
• Default Branch: master
• Last Pushed: 2022-07-20T09:39:26.000Z (over 2 years ago)
• Last Synced: 2024-10-14T20:21:25.621Z (23 days ago)
• Topics: keras, tensorflow2, wavelet-transform
• Language: Python
• Homepage:
• Size: 72.3 KB
• Stars: 10
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# cwtLayerKeras

  [![Generic badge](https://img.shields.io/badge/python-v3.6+-.svg)]() [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)

This project main goal is to produce a simple way to compute **CWT** (Continuous Wavelet Transformation) scalogram on signals :satellite: with keras functional API.

![Screenshot](cwtLayerKeras/cwt_performance.png)

The image above shows the performance of the layer on a TitanX GPU

## Installation

    pip install cwtLayerKeras

## Usage

```python

from cwtLayerKeras import Cwt as cwtt

# some code

def build_model():

    input_data = layers.Input(shape=(x_val.shape[-1],))

    cwt = cwtt(

        sample_count=x_val.shape[-1],

        scales_step=10,

        min_scale = 4,

        max_scale=224,

        output_size=(224, 224),

        depth_pad=2,

    )

    scalogram = cwt(input_data)

    base_model = EfficientNetB0(include_top=False, weights=None)(scalogram)

    x = layers.GlobalAveragePooling2D()(base_model)

    x = layers.BatchNormalization()(x)

    x = layers.Dropout(0.3)(x)

    predictions = layers.Dense(OUTPUT_SIZE, activation="softmax")(x)

    model = Model(inputs=input_data, outputs=predictions)

    model.summary()

    model.compile(

        optimizer="adam",

        loss="categorical_crossentropy",

        metrics=["accuracy"],

    )

    return model

```

## Parameters

Paramenter | Type | Description | Why

--- | --- | --- | ---

depth_pad | int | It allows to specify the number of padding channel that the layer need to add | some standard keras model (like efficientnet) need 3 channel depth to execute, instead of just the single one of the scalogram

max_scale|int|Highest wavelet's scale|

min_scale|int|Lowest wavelet's scale|

output_size|tuple|Size of the returned scalograms|

sample_count|int|# of Raman's shifts present in the tensor (in a tabular rapresentation,this number represent the amout of columns present in the table) |

scales_step|float|Change wavelet's scale by this range step|

trainable_kernels | bool | It make the wavelets that produce the scalogram trainable | interesting to see how/if the net optimizers change the wavelet kernel (research purpose)

wavelet|str|Wavelet kernel to use in the CWT. Now there are just "morl" (morlet), "mex_hat"/"rick" (mexican hat / ricker) |

More insightful informations can be found in the code's docs :wink:

## Disclaimer:

 :exclamation: This project was built for research purpose, so there could be some errors:exclamation:

 

 Feel free to open issues on those.

 

 Contributers are welcome :thumbsup: