app_model: Shiny app to fit a model or run a fitted model

In soundClass: Sound Classification Using Convolutional Neural Networks

Shiny app to fit a model or run a fitted model

Description

Shiny app to fit a model from training recordings or to run a fitted model to classify new recordings. This app consists of three GUIs, i.e. three main panels, accessible by the tabs at the top:

1. Create train data – create train data from recordings and their respective annotations database

2. Fit model – fit a model from training data

3. Run model – run a fitted model to classify new recordings

1. Create train data

This panel is used to create train data from recordings and their respective annotations database. The sidebar panel has the following buttons/boxes to input required user data:

• Choose folder – choose the folder containing the training recordings

• Choose database – choose the database with the annotations for the training recordings

• Time expanded – choose the correct time expansion factor, normally only used in recorders specifically intended for bat recordings. Can take the values "auto", 1 or 10. If the recording is in real time the value must be 1. If it's time expanded, the value 10 or "auto" can be selected. If "auto" is selected it is assumed that sampling rates < 50kHz corresponds to a value of 10 and sampling rates > 50kHz to corresponds to a value of 1

• Spectrogram parameters – different typologies of sound events require different parameters for computing the spectrograms. The more relevant are: size (in ms), which should be large enough to encompass the duration of the largest sound event in analysis (not only in the training data but also in novel recordings where the classifiers are to be applied) and moving window (in ms), that should be smaller for shorter sound events (to capture the quick changes in time) and larger for longer sound events (to avoid redundant information). The other parameters are more generalist and the same values can be used for different sound events, as they only change the definition of the images created. Please refer to spectro_calls documentation for further details

After entering the required information press the button "Create training data from labels" to generate the training data that will be used for fitting a model. This object is saved in the folder containing the training recordings with the name "train_data.RDATA".

2. Fit model

This panel is used to fit a model from training data. The sidebar panel has the following buttons/boxes to input required user data:

• Choose train data – the file "train_data.RDATA" created in the previous panel

• Choose model – a blank model to be fitted. A custom model is provided but must be copied to an external folder if it is to be used. The model path can be obtained by running the following line at the R console: system.file("model_architectures", "model_vgg_sequential.R", package="soundClass") and should be manually copied to a an external folder

• Model parameters – the train percentage indicates the percentage of data that is used to fit the model while the remaining are used for validation, batch size indicates the number of samples per gradient update, the learning rate indicates the degree of the gradient update, early stop indicates the maximum number of epochs without improvement allowed before training stops and epochs indicate the maximum number of epochs to train. Further information can be found in keras documentation https://keras.io/api/

The model is evaluated during fitting using the validation data. After completion, by reaching the maximum epochs or the early stopping parameters, the fitted model, the fitting log and the model metadata are saved to the folder containing the train data with file names: "fitted_model.hdf5", "fitted_model_log.csv" and "fitted_model_metadata.RDATA" respectively.

3. Run model

This panel is used to run a fitted model to classify new recordings. The sidebar panel has the following buttons/boxes to input required user data:

• Choose folder – choose the folder containing the recordings to be classified

• Choose model – a fitted model to be used for classification

• Choose metadata – the file containing the fitted model metadata

• Time expanded – choose the correct time expansion factor, normally only used in recorders specifically intended for bat recordings. Can take the values "auto", 1 or 10. If the recording is not time expanded the value must be 1. If it's time expanded, the value 10 or "auto" can be selected. If "auto" is selected it is assumed that sampling rates < 50kHz corresponds to a value of 10 and sampling rates > 50kHz to corresponds to a value of 1

• Output file – the name of the files to store the results of the classification

• Irrelevant – does the fitted model includes an irrelevant class?

• Export plots – should a spectrogram of the classified recordings be saved to disk?

The classification results are stored in a folder called "output", created inside the folder containing the recordings. They are stored in a database in sqlite3 format with all the relevant events detected and the respective probability of belonging to a given class. Additionally a file in the csv format is saved to disk, containing summary statistics per recording, i.e. the class with most events detected in each particular recording and the average frequency of maximum energy of the events detected.

Usage

app_model()

Value

Starts the shiny app, no return value.

Author(s)

Bruno Silva

soundClass documentation built on May 30, 2022, 1:07 a.m.