Last active
April 26, 2024 14:36
-
-
Save skjerns/8ce77afc6d79dbf21dbc9a0f1a9117d4 to your computer and use it in GitHub Desktop.
Extract 1 dimensional windows from ndarrays
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| import warnings | |
| import numpy as np | |
| from sklearn import feature_extraction | |
| def extract_windows(arr, sfreq, win_size, step_size, axis=-1): | |
| """extract 1d signal windows from an ndimensional array | |
| window_size and step_size are defined in terms of seconds | |
| this will extract a so called 'view' to the array, so the memory footprint | |
| is the same as the original array as no data is copied. The resulting | |
| views are therefore write protected to prevent accidental alteration. | |
| Parameters | |
| ---------- | |
| arr : np.ndarray | |
| input array | |
| sfreq : int | float | |
| sampling frequency. | |
| window_size : int | float | |
| window size in seconds. | |
| step_size : int | float | |
| window size in seconds. | |
| axis : int, optional | |
| axis that is denominating the time. The default is -1. | |
| Returns | |
| ------- | |
| windows : np.ndarray | |
| array with the extracted windows as the last two dimensions | |
| """ | |
| win_size_samples = win_size*sfreq | |
| step_size_samples = step_size*sfreq | |
| if np.round(win_size_samples)!=(win_size_samples:=int(win_size_samples)): | |
| rounded_length = win_size_samples/sfreq | |
| warnings.warn(f'{win_size=} s cannot accurately be represented with {sfreq=}, using {rounded_length:.3f} s') | |
| if np.round(step_size_samples)!=(step_size_samples:=int(step_size_samples)): | |
| rounded_length = step_size_samples/sfreq | |
| warnings.warn(f'{step_size=} s cannot accurately be represented with {sfreq=}, using {rounded_length:.3f} s') | |
| patch_shape = np.ones_like(arr.shape) | |
| extraction_step = np.ones_like(arr.shape) | |
| patch_shape[axis] = win_size_samples | |
| extraction_step[axis] = step_size_samples | |
| assert patch_shape[axis]<=arr.shape[axis], f'requested {win_size_samples=} > {arr.shape[axis]} of {axis=}' | |
| windows = feature_extraction.image._extract_patches(arr, patch_shape=patch_shape, | |
| extraction_step=extraction_step) | |
| # last but not least get rid of the singular dimensions | |
| new_shape = list(windows.shape[:arr.ndim]) + [x for x in windows.shape[arr.ndim:] if x>1] | |
| # there are some empty dimension | |
| windows = windows.reshape(new_shape) | |
| # make read-only to prevent accidental changes in views | |
| windows.flags.writeable=False | |
| return windows | |
| sfreq=750 | |
| # simulate data with 144 epochs, 306 channels and 4 seconds of data | |
| arr = np.random.rand(144, 306, sfreq*4) | |
| win_size = 0.5 # extract windows of 500 ms | |
| step_size = 0.25 # take new window every 250 ms | |
| axis=-1 # the time dimension is the last dimension | |
| wins = extract_windows(arr, sfreq, win_size, step_size, axis=axis) | |
| # wins.shape = (144, 306, 7, 500) -> 7 windows extracted for each channel for each epoch | |
| ###### | |
| #the same can be achieved with using sample values instead of seconds | |
| window = [1, 1, 500] # extract windows of size 500 samples i.e. 375 ms | |
| step = [1, 1, 250] # stride step size, take a window every 250 sample steps i.e. 187 ms | |
| wins2 = feature_extraction.image._extract_patches(arr, window, extraction_step=step) | |
| # wins2.shape = (144, 306, 7, 1, 1, 500) | |
| wins2 = wins2.squeeze() # call squeeze to get rid of the singular dimensions |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment