Updates

1.5.0

New key-word argument network_sum can be True, to compute network-summed correlation coefficients (default behavior, identical to previous versions) or False, to compute singe-channel correlation coefficients.

New key-word argument normalize can be 'short' or 'full'. 'short' computes a simplified correlation coefficient that assumes the signal in every sliding window has a mean of zero (initial and default implementation of FMF). 'full' computes the full correlation coefficient and is slower. NB: 'full' cannot be used with arch='cpu'.

arch can now be 'precise' in addition to 'cpu' or 'gpu'. 'precise' is a CPU implementation that does not use an optimized summation algorithm to speed up the calculation of the sum of the squared data. Thus, 'precise' is less fast than 'cpu' but does not lose in accuracy when large amplitudes are encountered in the data (which can sometimes happen with 'cpu').
The sum of the squared templates is computed only once at the beginning.
The station and component axes of the input arrays can be merged into a single axis of traces.

Fixed a bug in the computation of the sum of the squared data that occurred when some of the weights were equal to zero.