Version 0.8.0 (2021-12-07)

This is a minor version with new features, bug fixes, deprecations, and documentation improvements.

New Features

  • Added more efficient ways for creating the spline basis, and now solve penalized spline equations as a banded system rather than as a sparse system. Compared to version 0.7.0, spline.mixture_model, spline.irsqr, and morphological.mpspline are ~60-90% faster when numba is installed and ~10-70% faster without numba.

  • Made several calculations in spline.mixture_model more efficient, further reducing the time by ~60-70% compared to the timings above without numba. The total time reduction from version 0.7.0 for spline.mixture_model without numba is ~50-90%.

  • Added penalized spline versions of all Whittaker-smoothing-based algorithms (pspline_asls, pspline_iasls, pspline_airpls, pspline_arpls, pspline_drpls, pspline_iarpls, pspline_aspls, pspline_psalsa, and pspline_derpsalsa) to pybaselines.spline.

Bug Fixes

  • Was not multiplying the penalty in whittaker.iasls by lam_1.

  • The output weights for polynomial.quant_reg and polynomial.loess are now squared before returning since the square root of the weights are used internally.

  • The output weights and polynomial coefficients (if return_coef is True) for polynomial.loess are now sorted to match the original order of the input x-values.

  • The output weights for optimizers.optimize_extended_range are now truncated and sorted before returning to match the original order and length of the input x-values.

  • smooth.noise_median now works with a smooth_half_window value of 0 to give no smoothing.

Other Changes

  • Officially list Python 3.10 as supported.

  • pybaselines is now available to install using conda from the conda-forge channel.

  • Changed a factor in the weighting for whittaker.aspls to better match the implementation in literature.

  • Allow inputting x-values for all penalized spline functions rather than assuming evenly spaced measurements.

  • optimizers.adaptive_minmax now allows separate constrained_fraction and constrained_weight values for for the left and right edges.

  • The error raised by optimizers.collab_pls if the input data is not 2-dimensional is now more explicit.

Deprecations/Breaking Changes

  • No longer allow negative or array-like values for the penalty multipliers in Whittaker-smoothing-based functions, penalized spline functions, morphological.jbcd, or misc.beads. Array-like penalty values are technically valid; however, they change the symmetry of the banded linear system, so additional code will have to be added in a later version to ensure the setup is correct before re-allowing array-like values.

  • Deprecated passing keyword arguments to all functions in pybaselines.optimizers. Passing additional keyword arguments will raise an error starting in version 0.10.0 or 1.0.0, whichever comes first (the same deprecation for optimize_extended_range made in version 0.7.0 is also pushed back to 0.10.0 or 1.0.0).

  • For spline algorithms, the min and max x-values are now included as inner knots when creating the spline basis rather than counting them as the first outer knots. To match the number of knots from previous versions, the num_knots parameter should add 2 to the num_knots used in previous versions.

  • Formally deprecated pybaselines.window, which was replaced by pybaselines.smooth in version 0.6.0. pybaselines.window will be removed in version 1.0.

  • Removed optimize_window from pybaselines.morphological, which was deprecated in version 0.6.0

  • Removed the code for allowing array-like half_window or smooth_half_window values for morphological.rolling_ball, which was deprecated in version 0.7.0.


  • Added more examples to the documentation for fitting noisy data and exploring penalized spline parameters.

  • Added an introduction for the splines category in the algorithms section of the documentation.

Version 0.7.0 (2021-10-28)

This is a minor version with new features, bug fixes, deprecations, and documentation improvements.

Notice: beginning in version 0.8.0, a DeprecationWarning will be emitted when using any function from the pybaselines.window module. Use the pybaselines.smooth module instead.

New Features

  • Added the range independent algorithm (ria) to pybaselines.smooth, which extends the left and/or right edges, similar to optimizers.optimize_extended_range, and iteratively smooths until the area of the extended regions is recovered.

  • Added the joint baseline correction and denoising algorithm (jbcd) to pybaselines.morphological, which uses regularized least-squares fitting combined with morphological operations to simultaneously obtain the baseline and denoised signal.

  • Added the iterative polynomial smoothing algorithm (ipsa) to pybaselines.smooth, which iteratively smooths the input data using a second-order Savitzky–Golay filter.

  • Added the continuous wavelet transform baseline recognition algorithm (cwt_br) to pybaselines.classification, which uses a continuous wavelet transform to classify the baseline points and iterative polynomial fitting to create the baseline.

  • Added the fully automatic baseline correction algorithm (fabc) to pybaselines.classification, which is very similar to classification.dietrich, except that it uses a continuous wavelet transform to estimate the derivative and fits the baseline using Whittaker smoothing.

  • Added a min_length parameter to most classification algorithms, which allows discarding any values in the baseline mask where the number of consecutive points designated as baseline is less than min_length, making the algorithms more robust.

  • The threshold for polynomial.fastchrom can now be a Callable to allow the user to define their own thresholding functions based on the rolling standard deviation distribution.

  • Allow optimizers.optimize_extended_range to use spline (mixture_model, irsqr) and classification (dietrich, cwt_br, fabc) functions.

  • Allow optimizers.collab_pls to use spline functions (mixture_model, irsqr).

Bug Fixes

  • Increased the minimum scipy version to 1.0 in order to use the BLAS function gbmv (dot product of a banded matrix and vector) for misc.beads.

  • Use stable sorting when sorting the x-values for polynomial.loess and optimizers.optimize_extended_range to ensure that the sorting is correct.

  • Fixed an issue when specifying output with scipy.ndimage.uniform_filter1d in scipy versions before version 1.1.0.

  • Fixed an issue using dtype with numpy.arange in a numba jit wrapped function, which was not introduced until numba version 0.47.

  • Fixed an indexing error in spline.corner_cutting which would give an erroneous index at which the maximum area removal occurred.

  • Fixed an issue that occurred when inputting weights into spline.mixture_model.

  • If weights are input into optimizers.optimize_extended_range as keyword arguments, the weights are now correctly sorted to match the sorting of the x-values and padded to account for the added portions on the left and/or right edges before using in the fitting function.

  • Fixed the output of utils.padded_convolve when the kernel was even shaped (which never happens in actual application in pybaselines) or larger than the data.

  • Fixed an issue caused by using an extrapolate_window of 1 for utils.pad_edges, or an extrapolate_window of 0 or 1 for utils._get_edges (called by optimizers.optimize_extended_range).

Other Changes

  • Use scipy's expit function for whittaker.arpls and aspls, which does not emit the warning for exponential overflow. The warning was not needed since the overflow ultimately makes weights of 0 for the two functions.

  • Use np.gradient for the computed derivatives in derpsalsa and dietrich, which gives slightly less noisy derivatives than the finite difference used by np.diff.

  • Only sort x-values if they are given for polynomial.loess and optimizers.optimize_extended_range, which saves a little time otherwise.

  • Made whittaker.airpls error handling more robust in order to catch errors from the solvers as well, which should catch any errors not prevented by checking the residual's length.

  • Allow the mode for utils.pad_edges to be a callable padding function, matching numpy.pad's behavior.

  • Added tol_history to the output parameters of classification.dietrich.

  • Switched to using Scipy's convolve over Numpy's. Scipy's convolve can choose between the direct convolution, which is always used by Numpy, or an FFT based convolution, which is significantly faster for large arrays.

  • Added testing for the minimum supported versions of all dependencies to the project's continuous integration in order to ensure that the minimum stated dependencies actually work.

  • Allow specifying two separate extrapolate windows when padding using utils.pad_edges to allow better flexibility for fitting the edges.

Deprecations/Breaking Changes

  • Deprecated allowing passing additional keyword arguments to optimizers.optimize_extended_range since the pad_kwargs parameter is used by both the optimize_extended_range function and the internal functions it supports. Now, all keyword arguments should be placed in the method_kwargs dictionary. Passing additional keyword arguments will raise an error starting in version 0.9.0.

  • Deprecated allowing an array for the half_window or smooth_half_window parameters in morphological.rolling_ball. While the array-based moving min/max functions were valid, when combined for the morphological opening, the output would produce invalid results where the opening values were greater than the input data, which should not be allowed by the actual morphological opening. Using an array half_window will raise an error in version 0.8.0.


  • Added several new examples that explore different aspects of pybaselines.

  • Use sphinx-gallery to display the example programs' code and outputs within the documentation.

Version 0.6.0 (2021-09-09)

This is a minor version with new features, bug fixes, deprecations, and documentation improvements.

New Features

  • Added goldindec to pybaselines.polynomial, which uses a non-quadratic cost function with a shrinking threshold to fit the baseline.

  • Added the morphological penalized spline (mpspline) algorithm to pybaselines.morphological, which uses morphology to identify baseline points and then fits the points using a penalized spline.

  • Added the derivative peak-screening asymmetric least squares algorithm (derpsalsa) to pybaselines.whittaker, which includes additional weights based on the first and second derivatives of the data.

  • Added the fastchrom algorithm to pybaselines.classification, which identifies baseline points as where the rolling standard deviation is less than the specified threshold.

  • Added the module pybaselines.spline, which contains algorithms that use splines to create the baseline.

  • Added the mixture model algorithm (mixture_model) to pybaselines.spline, which uses a weighted penalized spline to fit the baseline, where weights are calculated based on the probability each point belongs to the noise.

  • Added iterative reweighted spline quantile regression (irsqr) to pybaselines.spline, which uses penalized splines and iterative reweighted least squares to perform quantile regression on the data.

  • Added the corner-cutting algorithm (corner_cutting) to pybaselines.spline, which iteratively removes corner points and then fits a quadratic Bezier spline with the remaining points.

Bug Fixes

  • Fixed an issue with utils.pad_edges when mode was "extrapolate" and extrapolate_window was 1.

Other Changes

  • Increased the minimum SciPy version to 0.17 in order to use bounds with scipy.optimize.curve_fit.

  • Changed the default extrapolate_window value in pybaselines.utils.pad_edges to the input window length, rather than 2 * window length + 1.

  • Slightly sped up pybaselines.optimizers.adaptive_minmax when poly_order is None by using the numpy array's min and max methods rather than the built-in functions.

Deprecations/Breaking Changes

  • Renamed pybaselines.window to pybaselines.smooth to make its usage more clear. Using pybaselines.window will still work for now, but will begin emitting a DeprecationWarning in a later version (maybe version 0.8 or 0.9) and will be removed shortly thereafter.

  • Removed the constant utils.PERMC_SPEC that was deprecated in version 0.4.1.

  • Deprecated the function pybaselines.morphological.optimize_window, which will be removed in version 0.8.0. Use pybaselines.utils.optimize_window instead.


  • Fixed the plot for morphological.mpls in the documentation.

  • Fixed the weighting formula for whittaker.arpls in the documentation.

  • Fixed a typo for the cost function in the docstring of misc.beads.

  • Updated the example program for all of the newly added algorithms.

Version 0.5.1 (2021-08-10)

This is a minor patch with bug fixes and minor changes.

Bug Fixes

  • Added classification to the main pybaselines namespace so that calling pybaselines.classification works correctly.

Other Changes

  • Changed the default tol for pybaselines.polynomial.quant_reg to 1e-6 to get better results.

  • Directly use the input eps value for pybaselines.polynomial.quant_reg rather than its square.

Version 0.5.0 (2021-08-02)

This is a minor version with new features, bug fixes, and deprecations.

New Features

  • Added quantile regression (quant_reg) to pybaselines.polynomial, which uses quantile regression to fit a polynomial to the baseline.

  • Added the top-hat transformation (tophat) to pybaselines.morphological, which estimates the baseline using the morphological opening.

  • Added the moving-window minimum value (mwmv) pybaselines.morphological, which estimates the baseline using the rolling minimum values.

  • Added the baseline estimation and denoising with sparsity (beads) method to pybaselines.misc, which decomposes the input data into baseline and pure, noise-free signal by modeling the baseline as a low pass filter and by considering the signal and its derivatives as sparse.

  • Added the module pybaselines.classification, which contains algorithms that classify baseline and/or peak segments to create the baseline.

  • Added Dietrich's classification method (dietrich) to pybaselines.classification, which classifies baseline points by analyzing the power spectrum of the data's derivative and then iteratively fits the points with a polynomial.

  • Added Golotvin's classification method (golotvin) to pybaselines.classification, which breaks the data into segments, uses the minimum standard deviation of all the segments to define the standard deviation of the entire data, and then classifies baseline points using that value.

  • Added the standard deviation distribution method (std_distribution) to pybaselines.classification, which classifies baseline segments by grouping the rolling standard deviation values into a distribution for the baseline and a distribution for the signal.

  • Added Numba as an optional dependency. Currently, the functions pybaselines.polynomial.loess, pybaselines.classification.std_distribution, and pybaselines.misc.beads are faster when Numba is installed.

  • When Numba is installed, the pybaselines.polynomial.loess calculation is done in parallel, which greatly improves the speed of the calculation.

  • The pybaselines.polynomial.loess function now takes a delta parameter, which will use linear interpolation rather than weighted least squares fitting for all but the last x-values that are less than delta from the last-fit x-value. Can significantly reduce calculation time.

  • All iterative methods now return an array of the calculated tolerance value for each iteration in the dictionary output, which should help to pick appropriate tol and max_iter values.

Bug Fixes

  • Added checks for airpls, drpls, and iarpls functions in pybaselines.whittaker to prevent nan or infinite weights in edge cases where too many iterations were done.

  • The baseline returned from polynomial algorithms was the second-to-last iteration's baseline, rather than the last iteration's. Now the returned baseline is the last iteration's.

  • Sort input weights and y0 (if use_original is True) for pybaselines.polynomial.loess after sorting the x-values, rather than leaving them unsorted.

Other Changes

  • Added a custom ParameterWarning for when a user-input parameter is valid but outside the recommended range and could cause issues with a calculation.

  • Changed the default conserve_memory value in polynomial.loess to True, since it is just as fast as False when Numba is installed and is safer.

  • pybaselines.optimizers.collab_pls now includes the parameters from each function call in the dictionary output as items in lists.

Deprecations/Breaking Changes

  • The key for the averaged weights for pybaselines.optimizers.collab_pls is now 'average_weights' to avoid clashing with the 'weights' key from the called function.


  • Most algorithms in the documentation now include several plots showing how the algorithm fits different types of baselines.

  • Added more in-depth explanations for all baseline correction algorithms.

Version 0.4.1 (2021-06-10)

This is a minor patch with new features, bug fixes, and pending deprecations.

New Features/Improvements

  • Switched to using banded solvers for all Whittaker-smoothing-based algorithms (all functions in pybaselines.whittaker as well as pybaselines.morphological.mpls), which reduced their computation time by ~60-85% compared to version 0.4.0.

  • Added pentapy as an optional dependency. All Whittaker-smoothing-based functions will use pentapy's solver, which is faster than SciPy's solve_banded and solveh_banded functions, if pentapy is installed and the system is pentadiagonal (diff_order is 2). All Whittaker functions with pentapy installed take ~80-95% less time compared to pybaselines version 0.4.0.

Bug Fixes

  • The alpha item in the dictionary output of whittaker.aspls is now the full alpha array rather than a single value.

  • The weighting for several Whittaker-smoothing-based functions was made more robust and less likely to create nan weights.

Other Changes

  • Increased the default max_iter for whittaker.aspls to 100.

Deprecations/Breaking Changes

  • The constant pybaselines.utils.PERMC_SPEC is no longer used. It will be removed in version 0.6.0.

Version 0.4.0 (2021-05-30)

This is a minor version with new features, bug fixes, and deprecations.

New Features/Improvements

  • Significantly reduced both the calculation time and memory usage of polynomial.loess. For example, getting the baseline for a dataset with 20,000 points now takes ~12 seconds and ~0.7 GB of memory compared to ~55 seconds and ~3 GB of memory in version 0.3.0.

  • Added a conserve_memory parameter to polynomial.loess that will recalculate the distance kernels each iteration, which is slower than the default but uses very little memory. For example, using loess with conserve_memory set to True on a dataset with 20,000 points takes ~18 seconds while using ~0 GB of memory.

  • Allow more user inputs for optimizers.optimize_extended_range to allow specifying the range of lam/poly_order values to test and to have more control over the added lines and Gaussians on the sides.

  • Added a constant called PERMC_SPEC (accessed from pybaselines.utils.PERMC_SPEC), which is used by SciPy's sparse solver when using Whittaker-smoothing-based algorithms. Changed the default value to "NATURAL", which reduced the computation time of all Whittaker-smoothing-based algorithms by ~5-35% compared to other permc_spec options on the tested system.

  • misc.interp_pts (formerly manual.linear_interp) now allows specifying any interpolation method supported by scipy.interpolate.interp1d, allowing for methods such as spline interpolation.

Bug Fixes

  • Fixed poly_order calculation for optimizers.adaptive_minmax when poly_order was a single item within a container.

  • Potential fix for namespace error with utils; accessing pybaselines.utils gave an attribute error in very specific envinronments, so changed the import order in pybaselines.__init__ to potentially fix it. Updated the quick start example in case the fix is not correct so that the example will still work.

  • Increased minimum NumPy version to 1.14 to use rcond=None with numpy.linalg.lstsq.

Other Changes

  • polynomial.loess now allows inputting weights, specifying a use_original keyword for thresholding to match the modpoly and imodpoly functions, and specifying a return_coef keyword to allow returning the polynomial coefficients for each x-value to recreate the fitted polynomial, to match all other polynomial functions.

  • Changed the default smooth_half_window value in window.noise_median, window.snip, and morphological.mormol to None, rather than being fixed values. Each function sets its default slightly different but still follows the behavior in previous versions, except for window.noise_median as noted below.

  • Changed default smooth_half_window value for window.noise_median to match specified half_window value rather than 1.

  • Changed default sigma value for window.noise_median to scale with the specified smooth_half_window, rather than being a fixed value.

Deprecations/Breaking Changes

  • Renamed pybaselines.manual to pybaselines.misc to allow for adding any future miscellaneous algorithms that will not fit elsewhere.

  • Renamed the manual.linear_interp function to misc.interp_pts to reflect its more general interpolation usage.

  • The parameter dictionary returned from Whittaker-smoothing-based functions no longer includes 'roughness' and 'fidelity' values since the values were not used elsewhere.

Version 0.3.0 (2021-04-29)

This is a minor version with new features, bug fixes, deprecations, and documentation improvements.

New Features/Improvements

  • Added the small-window moving average (swima) baseline to pybaselines.window, which iteratively smooths the data with a moving average to eliminate peaks and obtain the baseline.

  • Added the rolling_ball function to pybaselines.morphological, which applies a minimum and then maximum moving window, and subsequently smooths the result, giving a baseline that resembles rolling a ball across the data. Also allows giving an array of half-window values to allow the ball to change size as it moves across the data.

  • Added the adaptive_minmax algorithm to pybaselines.optimizers, which uses the modpoly or imodpoly functions and performs polynomial fits with two different orders and two different weighting schemes and then uses the maximum values of all the baselines.

  • Added the Peaked Signal's Asymmetric Least Squares Algorithm (psalsa) function to pybaselines.whittaker, which uses exponentially decaying weighting to better fit noisy data.

  • The imodpoly and loess functions in pybaselines.polynomial now use num_std to specify the number of standard deviations to use when thresholding.

  • The pybaselines.polynomial.penalized_poly function now allows weights to be used. Also made the default threshold value scale with the data better.

  • Added higher order filters for pybaselines.window.snip to allow for more complicated baselines. Also allow inputting a sequence of ints for max_half_window to better fit asymmetric peaks.

Bug Fixes

  • Fixed a bug that would not allow even morphological half windows, since it is not needed for the half windows, only the full windows.

  • Fixed the thresholding for pybaselines.polynomial.imodpoly, which was incorrectly not adding the standard deviation to the baseline when thresholding.

  • Fixed weighting for pybaselines.whittaker.airpls so that weights no longer get values greater than 1.

  • Removed the append and prepend keywords for np.diff in the pybaselines.morphological.mpls function, since the keywords were not added until numpy version 1.16, which is higher than the minimum stated version for pybaselines.

Other Changes

  • Allow utils.pad_edges to work with a pad_length of 0 (no padding).

  • Added a 'min_half_window' parameter for pybaselines.morphological.optimize_window so that small window sizes can be skipped to speed up the calculation.

  • Changed the default method from 'aspls' to 'asls' for optimizers.optimize_extended_range.

Deprecations/Breaking Changes

  • Removed the 'smooth' keyword argument for pybaselines.window.snip. Smoothing is now performed if the given smooth half window is greater than 0.

  • pybaselines.polynomial.loess no longer has an include_stdev keyword argument. Equivalent behavior can be obtained by setting num_std to 0.


  • Updated the documentation to include simple explanations for some techniques.

Version 0.2.0 (2021-04-02)

This is a minor version with new features, bug fixes, deprecations, and documentation improvements.

New Features/Improvements

  • Added the morphological and mollified (mormol) function to pybaselines.morphological, which uses a combination of morphology for baseline estimation and mollification for smoothing.

  • Added the loess function to pybaselines.polynomial, which does local robust polynomial fitting. Allows using symmetric or asymmetric weighting, or using thresholding, similar to the modpoly and imodpoly functions.

  • Added the penalized_poly function to pybaselines.polynomial, which fits a polynomial baseline using a non-quadratic cost function. The non-quadratic cost functions include huber, truncated-quadratic, and indec, and can be either symmetric or asymmetric.

  • Added options for padding data when doing convolution or window-based operations to reduce edge effects and give better results.

Bug Fixes

  • Fixed the mollification kernel used for the morphological.iamor (now amormol) function.

  • Fixed a miscalculation with the weighting for whittaker.aspls.

Other Changes

  • Slightly sped up several functions in by precomputing terms.

  • Added tests for all baseline algorithms

Deprecations/Breaking Changes

  • Renamed morphology.iamor to morphology.amormol (averaging morphological and mollified baseline) to make it more clear that mormol and amormol are similar methods.

  • Renamed to, to be more specific, since other techniques also use penalized least squares for polynomial fitting.


  • Updated the example program to match the changes to pybaselines.

  • Setup initial documentation.

Version 0.1.0 (2021-03-22)

  • Initial release on PyPI.