Visualizing data
beetles viz --debug_data_path <output of beetles infer>
The directory at --debug_data_path needs to contain five files
(produced by beetles infer --debug <viz_path>): classifications.csv,
raw_spectrogram.pkl, hmm_predictions.pkl, median_predictions.pkl, iqrs.pkl.
This command displays an matplotlib window with a slider to allow interactive visualization of ensemble predictions. Click or slide on a region of the slider to move the visualization window.
The app displays the raw spectrogram as the top image, along with annotations
for beginning and ends of different chirps. Background labels aren't shown.
The second image shows the median predictions as probabilities, the uncertainties associated with each prediction (labeled iqrs, for inter-quartile ranges), the class prediction of the ensemble (labeled median argmax), and the predictions post-hmm. The predictions post-hmm are used to create the vertical lines in the spectrogram plot.
The neural networks only predict three classes (A, B, and background) and as such we map each prediction to an RGB code. This means a prediction of A (output of the ensemble in this case is [1, 0, 0]) gets visualized as red. Accordingly, bright red, green, or blue in the first row means the ensemble was very confident that the chirp was A, B, or background, respectively.
The visualization of the uncertainty (iqr) is the same. Black means uncertainty was close to 0 for all classes (an output of [0, 0, 0]), and brighter colors mean more uncertainty. For example, cyan is encoded in RGB as [0, 1, 1]. This color means that the ensemble had high uncertainty between classes B and background (indices 1 and 2). In the image above, cyan often appears at the transition between B chirps and background, meaning that different members of the ensemble learned slightly different definitions of background and B chirp.
When trying to interpret a new color (yellow in the uncertainty row, for example), just find the RGB vector that describes it.