boxsers

---

Introduces BoxSERS, a complete and ready-to-use python library for the application of data augmentation, dimensional reduction, spectral correction, machine learning and other methods specially designed and adapted for vibrational spectra(Raman,FTIR, SERS, etc.).

Table of contents

• BoxSERS Installation
• Requirements
• Included Features
  • Module misc_tools
  • Module visual_tools
  • Module preprocessing
  • Module data augmentation
  • Dimensional Reduction
  • Unsupervised Machine Learning
  • Supervised Machine Learning
  • Module validation_metrics
• License

BoxSERS Installation

From PypY

pip install boxsers

From Github

pip install git+https://github.com/ALebrun-108/BoxSERS.git

Requirements

Listed below are the main modules needed to operate the codes:

• Sklearn
• Scipy
• Numpy
• Pandas
• Matplotlib
• Tensor flow (GPU or CPU)

Labels associated to spectra can be in one of the following three forms:

Label Type	Examples
Text	Cholic, Deoxycholic, Lithocholic, ...
Integer	0, 3, 1 , ...
Binary	[1 0 0 0], [0 0 0 1], [0 1 0 0], ...

Included Features

---

Module boxsers.misc_tools

This module provides functions for a variety of utilities.

• data_split : Randomly splits an initial set of spectra into two new subsets named in this function: subset A and subset B.

• load_rruff : Export a subset of Raman spectra from the RRUFF database in the form of three related lists containing Raman shifts, intensities and mineral names.

Module boxsers.visual_tools

This module provides different tools to visualize vibrational spectra quickly.

• spectro_plot : Returns a plot with the selected spectrum(s)

• random_plot : Plot a number of randomly selected spectra from a set of spectra.

• distribution_plot : Return a bar plot that represents the distributions of spectra for each classes in a given set of spectra

# Code example:
from boxsers.misc_tools import data_split
from boxsers.visual_tools import spectro_plot, random_plot, distribution_plot

wn = 3 
spec =5 

# randomly splits the spectra(spec) and the labels(lab) into test and training subsets.
(spec_train, spec_test, lab_train, lab_test) = data_split(wn, spec , b_size=0.4)  
# resulting train|test set proportions = 0.6|0.4

# plots the classes distribution within the training set.
distribution_plot(lab_train, title='Train set distribution')

# spectra array = spec, raman shift column = wn
random_plot(wn, spec, random_spectra=4)  # plots 4 randomly selected spectra
spectro_plot(wn, spec[0], spec[2])  # plots first and third spectra

Module boxsers.preprocessing

This module provides multiple functions to preprocess vibrational spectra. These features improve spectrum quality and can improve performance for machine learning applications.

• baseline_substraction : Subtracts the baseline signal from the spectrum(s) using Asymmetric Least Squares estimation.

• intensity_normalization : Normalizes the spectrum(s) using one of the available norms in this function.

• savgol_smoothing : Smoothes the spectrum(s) using a Savitzky-Golay polynomial filter.

• spectral_cut : Subtracts or sets to zero a delimited spectral region of the spectrum(s)

• spline_interpolation : Performs a one-dimensional interpolation spline on the spectra to reproduce them with a new x-axis.

# Code example:
import numpy as np
from boxsers.preprocessing import als_baseline_cor, spectral_cut, spectral_normalization, spline_interpolation

# interpolates with splines the spectra and converts them to a new raman shift range(new_wn)
new_wn = np.linspace(500, 3000, 1000)
spec_cor = spline_interpolation(spec, wn, new_wn)
# removes the baseline signal measured with the als method 
(spec_cor, baseline) = als_baseline_cor(spec, lam=1e4, p=0.001, niter=10)
# normalizes each spectrum individually so that the maximum value equals one and the minimum value zero 
spec_cor = spectral_normalization(spec)
# removes part of the spectra delimited by the Raman shift values wn_start and wn_end 
spec_cor, wn_cor = spectral_cut(spec, wn, wn_start, wn_end)

Module boxsers.data_augmentation

This module provides several data augmentation methods that generate new spectra by adding different variations to existing spectra.

• aug_mixup : Randomly generates new spectra by mixing together several spectra with a Dirichlet probability distribution.

• aug_noise : Randomly generates new spectra with Gaussian noise added.

• aug_multiplier : Randomly generates new spectra with multiplicative factors applied.

• aug_ioffset : Randomly generates new spectra shifted in intensity.

• aug_xshift : Randomly generates new spectra shifted in wavelength.

• aug_linslope : Randomly generates new spectra with additional linear slopes

# Code example:

from boxsers.data_augmentation import aug_mixup, aug_noise

spectra_nse, label_nse  = aug_noise(spec, lab, snr=10)
spectra_mult, label_mult = aug_multiplier(spectra, labels, 0.15,)
spectro_plot(wn, spec, spec_nse, spec_mult_sup, spec_mult_inf, legend=legend)

spec_nse, lab_nse = aug_noise(spec, lab, param_nse, quantity=2, mode='random')
spec_mul, lab_mul = aug_multiplier(spec, lab, mult_lim, quantity=2, mode='random')

# stacks all generated spectra and originals in a single array
spec_aug = np.vstack((x, spec_nse, spec_mul))
lab_aug = np.vstack((lab, lab_nse, lab_mul))

# spectra and labels are randomly mixed
x_aug, y_aug = shuffle(x_aug, y_aug)

Module boxsers.dimension_reduction

This module provides different techniques to perform dimensionality reduction of vibrational spectra.

• SpectroPCA: Returns a plot with the selected spectrum(s)

• SpectroPCA : Plot a number of randomly selected spectra from a set of spectra.

• distribution_plot : Return a bar plot that represents the distributions of spectra for each classes in a given set of spectra

Dimensional Reduction

# Code example:

from boxsers.machine_learning.dimension_reduction import SpectroPCA, SpectroICA

pca_model = SpectroPCA(n_comp=50)
pca_model.fit_model(spec_train)
pca_model.scatter_plot(spec_test, spec_test, targets=classnames, component_x=1, component_y=2)
pca_model.component_plot(wn, component=2)
spec_pca = pca_model.transform_spectra(spec_test)

Unsupervised Machine Learning

# Code example:

from boxsers.machine_learning import SpectroGmixture, SpectroKmeans

kmeans_model = SpectroKmeans(n_cluster=5)
kmeans_model.fit_model(spec_train)
kmeans_model.scatter_plot(spec_test)

Supervised Machine Learning

• Convolutional Neural Networt (3 x Convolutional layer 1D , 2 x Dense layer)

from boxsers.pca_model import SpectroPCA, SpectroFA, SpectroICA

pca_model = SpectroICA(n_comp=50)
pca_model.fit_model(x_train)
pca_model.scatter_plot(x_test, y_test, targets=classnames, comp_x=1, comp_y=2)
pca_model.pca_component(Wn, 2)
x_pca = pca_model.transform_spectra(x_train)

Module validation_metrics

This module provides different tools to evaluate the quality of a model’s predictions.

• cf_matrix : Returns a confusion matrix (built with scikit-learn) generated on a given set of spectra.

• clf_report : Returns a classification report generated from a given set of spectra