""" This repo is the raw code used in the paper "Deep neural networks to recover unknown physical parameters from oscillating time series" (https://doi.org/10.1371/journal.pone.0268439).
Architecture used in the paper: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0268439
The code generate a huge amount of frequency-modulated, amplitude-modulated and pure decaying sinewave. To which gaussian noise, spurious events and drifts are added. A custom DNN architecture is then produced as a base model. The training task is to reproduce denoised time-series, in addition to an approximation of the latent parameters (frequency, amplitude, phase, noise level, decay rate etc). After training, a series of experiments are run to evaluate performance and study possible applications of the models:
-1 raw performance (denoising and regression)
-2 performance vs a non-linear least-square fit (with true latent parameters as initial guesses)
-3 performance for a multi-model trained DNN
-4 application of the DNN to estimate initial guesses for a fit
-5 an interesting discussion on hyperparamters is also done
What comes out: 1- After training, the DNN outperform least-square fit (with true latent parameters as initial guesses) in almost all cases.
2- When used as a preprocessing tool to informe least-square fit initial guesses, the outcome is always the optimal solution. This remove the need for initial guesses exploration when fitting a large number of signals.
3- The DNN is able to deal with multiple models at once, which is obviously never the case for a least-square fit. This is espcially usefull in processing unknown or varying signal-models.
This code is currently uncommented. How to proceed with the current version of the repo:
1- Generate datasets: Run all scripts in ./data
2- Instanciate DNN model: Run script in ./models
3- Train models: Run all scripts in ./trainModels /!\ These scripts are path dependent: look at the generated-folders names in ./models and ./data and adjust for yourself. /!\ Any decent training will take ~12/24h on a 1080GTX
4- Evaluate models and run experiments: Run all scripts in ./experiments
"""
DNN architecture output example on a noisy amplitude-modulated sinewave.
"Performance comparison of the specialized DNN (trained on AM-sine waves, tasked to denoise signals and recover all latent parameters of AM-sine waves) and of the partial DNN (trained on monochromatic, AM- and FM-sine waves, tasked to denoise signals and recover the carrier frequency, phase, coherence time and noise level only).
Top: Randomly selected example of noisy input AM-signal, alongside specialized- and partial-DNN denoised and latent predictions. Bottom: Individual latent parameters and signal denoising root mean squared error (RMSE), averaged over the whole AM-sinewave test set (100′000 samples) for both DNNs."