This function processes sound data from a specified directory, performs transfer learning using a pre-trained deep learning model, and saves the results.
deploy_CNN_multi(
output_folder,
output_folder_selections,
output_folder_wav,
top_model_path,
path_to_files,
windowlength = 512,
detect_pattern = NA,
architecture,
clip_duration = 12,
hop_size = 6,
downsample_rate = 16000,
threshold = 0.1,
save_wav = TRUE,
class_names = c("female.gibbon", "hornbill.helmeted", "hornbill.rhino", "long.argus",
"noise"),
noise_category = "noise",
min_freq_khz = 0.4,
max_freq_khz = 2,
single_class = TRUE,
single_class_category = "female.gibbon",
for_prrec = TRUE
)A character string specifying the path to the output folder where the results will be saved.
A character string specifying the path to the folder where selection tables will be saved.
A character string specifying the path to the folder where extracted WAV files will be saved.
A character string specifying the path to the pre-trained top model for classification.
A character string specifying the path to the directory or list containing sound files to process.
window length for input into 'spectro' function from seewave. Deafults to 512.
Pattern in sound file to detect for subset.
User specified: 'alexnet', 'vgg16', 'vgg19', 'resnet18', 'resnet50', or 'resnet152'
The duration of each sound clip in seconds.
The hop size for splitting the sound clips.
The downsample rate for audio in Hz, set to 'NA' if no downsampling is required.
The threshold for audio detection.
A logical value indicating whether to save the extracted sound clips as WAV files.
A character vector containing the unique classes for training the model.
A character string specifying the noise category for exclusion.
The minimum frequency in kHz for spectrogram visualization.
The maximum frequency in kHz for spectrogram visualization.
A logical value indicating whether to process only a single class. For now 'TRUE' is only option.
A character string specifying the single class category when 'single_class' is set to TRUE.
Whether to output all detections to create a PR curve.
Returns spectrogram images, wav files (if specified), and Raven selection tables for each sound file in specified output directory.
This function processes sound data from a directory, extracts sound clips, converts them to images, performs image classification using a pre-trained deep learning model, and saves the results including selection tables and image and audio files.
This function takes a model trained using the 'train_CNN_multi' function.
{
# Load data
data("TempBinWav")
# Create necessary directories
dir.create(file.path(tempdir(), "MultiDir", "Wav"), recursive = TRUE, showWarnings = FALSE)
# Write to temp directory
writeWave(TempBinWav, filename = file.path(tempdir(), "MultiDir", "Wav", "TempBinWav.wav"))
# Set model directory
trained_models_dir <- system.file("extdata", "trainedresnetmulti", package = "gibbonNetR")
# Specify model path
ModelPath <- list.files(trained_models_dir, full.names = TRUE)
# Deploy trained model over sound files
deploy_CNN_multi(
clip_duration = 12,
architecture = "resnet18",
output_folder = file.path(tempdir(), "MultiDir", "Results", "Images"),
output_folder_selections = file.path(tempdir(), "MultiDir", "Results", "Selections"),
output_folder_wav = file.path(tempdir(), "MultiDir", "Results", "Wavs"),
detect_pattern = NA,
top_model_path = ModelPath,
path_to_files = file.path(tempdir(), "MultiDir", "Wav"),
downsample_rate = "NA",
save_wav = FALSE,
class_names = c("female.gibbon", "hornbill.helmeted", "hornbill.rhino", "long.argus", "noise"),
noise_category = "noise",
single_class = FALSE,
threshold = 0.25,
max_freq_khz = 2
)
}
#> Warning: '/var/folders/1s/x8xb37tj45j86tn_stc4v44w0000gn/T//RtmpScZf01/MultiDir/Results/Images' already exists
#> Warning: '/var/folders/1s/x8xb37tj45j86tn_stc4v44w0000gn/T//RtmpScZf01/MultiDir/Results/Selections' already exists
#> Warning: '/var/folders/1s/x8xb37tj45j86tn_stc4v44w0000gn/T//RtmpScZf01/MultiDir/Results/Wavs' already exists
#> 1 out of 1
#> TempBinWav.wav
#> saving sound clips
#> Creating images 1 start time clips
#> Classifying images using Top Model
#> Warning: Some torch operators might not yet be implemented for the MPS device. A
#> temporary fix is to set the `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a
#> fall back for those operators:
#> ℹ Add `PYTORCH_ENABLE_MPS_FALLBACK=1` to your `.Renviron` file, for example use
#> `usethis::edit_r_environ()`.
#> ✖ Using `Sys.setenv()` doesn't work because the env var must be set before R
#> starts.
#> 2222
#> Saving output
#> Saving output to /var/folders/1s/x8xb37tj45j86tn_stc4v44w0000gn/T//RtmpScZf01/MultiDir/Results/Images/___TopModel_.jpg
#>
#> NANANANANANANANANANA
#> Saving Selection Table No Detections
#> 0.455532073974609