See recent release notes on the github repository
- pushing pyopencl-arrays caused error (thanks to @psobolewskiPhD for reporting)
combine_labelserode_connected_labelsexclude_labels_with_average_values_out_of_rangeexclude_labels_with_average_values_with_rangesubtract_labels
- Support ellipsis
a[1:, ...]when accessing GPU-backed images (thanks for reporting the issue @tlambert03) - Solved issue in deskew_x (thanks to @pr4deepr)
- Due to changes in how indexing arrays are handled, it is possible that in special cases, numpy arrays are returned where former versions of pyclesperanto_prototype raised an error. Example: An array-access that results in a 1d array will return a numpy array.
reduce_labels_to_label_edges
deskew_yproduced wrong results in case voxel_size_z was != 0.3. (Thanks to @DrLachie for reporting and fixing)
- The parameter 'destintation' in
absolute()was renamed todestination - The parameter 'input' of the following functions were renamed to 'source':
bottom_hat_boxbottom_hat_spherecreate_pointlist_from_labelmapcreate_vector_from_labelmapcreate_2d_xycreate_2d_yxcreate_2d_zycreate_2d_yzcreate_2d_zxcreate_2d_xzcropdifference_of_gaussiandivide_by_gaussian_backgroundexclude_labels_outside_size_rangeexclude_large_labelsexclude_small_labelslaplace_boxlocal_maximum_touching_neighbor_count_maplabel_mean_intensity_maplocal_mean_touching_neighbor_count_maplocal_median_touching_neighbor_count_maplocal_minimum_touching_neighbor_count_maplabel_pixel_count_maplocal_standard_deviation_touching_neighbor_count_mapnonzero_maximum_boxnonzero_maximum_diamondnonzero_minimum_boxnonzero_minimum_diamondonlyzero_overwrite_maximum_boxonlyzero_overwrite_maximum_diamondpower_imagesreduce_stackrelabel_sequentialreplace_intensitiesreplace_intensityset_non_zero_pixels_to_pixel_indexstatistics_of_background_and_labelled_pixelssubtract_gaussian_backgroundsubtract_image_from_scalarthreshold_otsutop_hat_boxtop_hat_spheretranspose_xytranspose_xztranspose_yztouching_neighbor_count_map
- optional parameters are now marked as such in the docstrings
- Deskewing produced wrong results in case voxel_size_z was != 0.3. This release contains a temporary workaround for fixing this. The workaround leads to higher memory consumption, longer compute time and tiny interpolation artefacts. A future release will fix those issues. (Thanks to @DrLachie for reporting and also for demonstrating the potential workaround)
- Support for pushing XArrays (thanks to @psobolewskiPhD for reporting a related issue)
- Slicing works now in Y and X, not just in Z. E.g.: x = y[:, 5, 2:7]
range()function for fancy indexingcopy_vertical_slice()similar tocopy_slice()function for fancy indexingcopy_horizontal_slice()similar tocopy_slice()function for fancy indexingeroded_otsu_labelinghas been added as potentially better alternative to Voronoi-Otsu-Labeling (Thanks to Jan Brocher @biovoxxel for inspiration and support)gauss_otsu_labelinghas been added as simpler and faster alternative to Voronoi-Otsu-Labelingclosing_labelsandopening_labelshave been added for post-processing label imagesclosing_/opening_box/spherehave been added for background removal and binary image post-processingalphaandcontinue_drawingparameters were added to theimshowfunction to enable overlaying intensity images and label images
maximum_boxandmaximum_spherehad wrong parameter type annotations which were corrected. Functionality remains unchanged.- The new slicing techniques above are intended to improve interoperability with numpy. The changes may lead to old workflows fail.
- The default colormap for
imshowhas been changed to 'Greys_r'
- The
color_mapparameter for imshow has been deprecated. Usecolormapinstead.
reduce_labels_to_centroidsmethod addedartificial_objects_2dmethod added, for providing a synthetic datasethessian_eigenvaluesfilter (yet with multiple outputs, based on code by Matthias Arzt @maarzt)small_hessian_eigenvaluefilter addedlarge_hessian_eigenvaluefilter added- Added
auto_sizeparameter to the functionsscale,rotate,rigid_transformandaffine_transformwhich will automatically determine the size of the destination image if it is not provided. With this, no pixels will be transformed out of the image anymore (Thanks to Pradeep Rajasekhar @pr4deepr). - Added
shear_in_?_planeaffine transforms (Thanks to Pradeep Rajasekhar @pr4deepr). - Added
deskew_xanddeskew_yoperation (Thanks to Pradeep Rajasekhar @pr4deepr). rigid_transformis now also available in the assistant.draw_distance_mesh_between_proximal_labelsAffineTransform3D.rotate_around_x_axisconvenience function was addedAffineTransform3D.rotate_around_y_axisconvenience function was addedAffineTransform3D.rotate_around_z_axisconvenience function was added
- The example image data "RBC_tiny" for deskewing was kindly provided by Ms Cindy Evelyn and Dr Niall Geoghegan and who used lattice light sheet microscopy (Zeiss) to capture red blood cells and investigate the malaria parasite invasion process.
- The example image data "45_CamA_ch0_stack0006_488nm_0016799msec_0013438399msecAbs-1.tif" kindly provided by Ms Cindy Evelyn , Dr Niall Geoghegan and Dr. Kelly Rogers from the Centre for Dynamic Imaging, WEHI, Melbourne, Australia who used lattice light sheet microscopy to capture red blood cells and study the malaria parasite invasion process.
threshold_otsudelivered different results compared to CLIJ and scikit-image, in case the minimum intensity in the input image was different from zero. (Thanks to Christophe Avenel @cavenel for reporting and fixing this bug!)touch_matrix_to_meshanddistance_matrix_to_meshcrashed when no destination image was provided.
- Performance optimization of the
pushcommand (Thanks to Elissavet Sandaltzopoulou @elsandal for implementing this) - The colormap / lookup table of the label-maps used in the
imshowfunction is no longer random. It will now allow reproducible visualizations.
average_distance_of_n_closest_neighbors_mapwas renamed toaverage_distance_of_n_nearest_neighbors_map. An alias for backwards compatibility was set. This alias will be removed in a later release.
- Removed support for float64 (double) pixel types. Float32 will be used instead.
- If
Imagetype annotated parameters of any clesperanto function are passed as tuples or lists, they are automatically converted usingnp.asarray()and pushed to the GPU. - Almost all functions have annotated return types now. While this should not affect functionality, code-parsers may perceive the API differently with this change.
- The bugfix in
threshold_otsumight change segmentation results, also when using related operations such asvoronoi_otsu_labeling.
New functions for processing and comparing label images:
average_distance_of_n_shortest_distancesaverage_distance_to_n_nearest_other_labels_mapproximal_other_labels_countproximal_other_labels_count_mapexclude_labels_with_values_equal_to_constantexclude_labels_with_values_not_equal_to_constantexclude_labels_with_map_values_equal_to_constantexclude_labels_with_map_values_not_equal_to_constantexclude_labels_with_map_values_out_of_rangeexclude_labels_with_map_values_within_range
New API for affine transforms
transform.shear_in_x_planetransform.shear_in_y_planetransform.shear_in_z_plane- The
transformparameter can now also be a string containing these definitions for transforms (similar to the CLIJ2 affine transforms API):- -center: translate the coordinate origin to the center of the image
- center: translate the coordinate origin back to the initial origin
- rotate=[angle]: rotate in Z-plane (around Z-axis) by the given angle in degrees
- rotate_X=[angle]: rotate in X-plane (around X-axis) by the given angle in degrees
- rotate_Y=[angle]: rotate in Y-plane (around Y-axis) by the given angle in degrees
- rotate_Z=[angle]: rotate in Z-plane (around Z-axis) by the given angle in degrees
- scale=[factor]: isotropic scaling according to given zoom factor
- scale_X=[factor]: scaling along X-axis according to given zoom factor
- scale_Y=[factor]: scaling along Y-axis according to given zoom factor
- scale_Z=[factor]: scaling along Z-axis according to given zoom factor
- shear_in_Z_plane_along_X=[angle]: shearing along X-axis in Z plane according to given angle in degrees
- shear_in_Z_plane_along_Y=[angle]: shearing along Y-axis in Z plane according to given angle in degrees
- shear_in_Y_plane_along_X=[angle]: shearing along X-axis in Y plane according to given angle in degrees
- shear_in_Y_plane_along_Z=[angle]: shearing along Z-axis in Y plane according to given angle in degrees
- shear_in_X_plane_along_Y=[angle]: shearing along Y-axis in X plane according to given angle in degrees
- shear_in_X_plane_along_Z=[angle]: shearing along Z-axis in X plane according to given angle in degrees
- translate_X=[distance]: translate along X-axis by distance given in pixels
- translate_Y=[distance]: translate along Y-axis by distance given in pixels
- translate_Z=[distance]: translate along Z-axis by distance given in pixels
generate_distance_maphad wrong output shape in case two centroid lists with different sizes were given
generate_distance_matrixhad a typo in the parameter "coordindate_list1": it's now "coordinate_list1".
- Draw mesh to nearest neighbors drawing was incorrect in 3D
- For improved code generation, pyclesperanto_prototype as now a
__common_alias__attribute.
- Interoperability with igraph and networkx
n_nearest_labels_to_igraphn_nearest_labels_to_networkxproximal_labels_to_igraphproximal_labels_to_networkxto_igraphto_networkxtouching_labels_to_igraphtouching_labels_to_networkx
- Functions made available in the napari-pyclesperanto-assistant
combine_verticallycombine_horizontallyconcatentate_stacksrelabel_sequential
- added
map_arrayalias toreplace_intensities; scikit-image users may be used to it - obfuscate pyclesperanto's internal structure. The
__module__of all functions is now the core package. Hence, one can no longer know from which_tiera function is by printing out the module. This is supposed to increase the quality of generated code.
- Deployment was broken. Version 0.11.0 may not have worked under any circumstance.
- This version introduces an experimental
cupybackend. Install cupy and runcle.select_device('cupy')to test it. This feature is experimental yet, please use it with care. These functions are known to cause issues with the cupy backend:- affine_transform (linear interpolation not supported),
- apply_vector_field (linear interpolation not supported),
- create(uint64), create(int32), create(int64),
- resample (linear interpolation not supported),
- scale (linear interpolation not supported),
- translate (linear interpolation not supported).
- The following parameters of
execute()are deprecated and will disappear with version 0.12.0.image_size_independent_kernel_compilation(will always be True in the future),prog(will no longer be supported, use pyopencl for those kinds of features) anddevice(will no longer be supported, use pyopencl instead.
The next version, 0.12.0 will bring larger internal changes, especially in the context of tier0.
If you used private functions such as in _tier0._pycl, please get in touch, e.g. via github issue.
- deployment switched from setup.py to setup.cfg
standard_deviation_box/standard_deviation_spherefilters (Thanks to Lucrezia Ferme @lcferme and Allyson Ryan @allysonryan for suggesting)
- Kernels were crashing when calling
y = x * x.max()in casexwas a GPU-backed image of type uint16 (and others).
imshowhas a new optional parameter:colorbar- added
label_nonzero_pixel_count_map - added
label_nonzero_pixel_count_ratio_map - added
label_overlap_count_map
imshowuses the same colormap when visualizing label images
- imshow(numpy array) crashed
- scikit-image regionprops crashed when passing a pyclesperanto labels image
- It is now possible to read and write pixels in images using the syntax
image[([z1,z2], [y1,y2], [x1,x2])] = new_value
- When importing
pyclesperanto_prototype, some functions inpyopenclstopped working (see #130) (Thanks to Peter Sobolewski for reporting and Talley Lambert for support while fixing) - Operations crashed when a dask array was handed over as image parameter. That can happen in napari for example when opening CZI files.
- z-position projections automatically created output images had a wrong size (x and y switched).
- Instead of manipulating internal sturctures of pyopencl (in particular
cl.array.Array), we now ship an own image class,OCLArray. Thus, it is possible that operators directly applied to images produced by pyclesperanto don't work anymore. If you experience any issues, please report them.
relabel_sequential, similar to scikit-image (formely known asclose_index_gaps_in_label_maps)erode_labelsextend_labels_with_maximum_radiushas been renamed todilate_labels, an alias to the old method is kept as wellexclude_small_labelsas shortcut / convenience function toexclude_labels_out_of_size_rangeexclude_large_labelsas shortcut / convenience function toexclude_labels_out_of_size_rangesearch_operation_namesfor searching in the clesperanto API
- Switching GPUs using
select_devicemade operations fail that were executed before switching (#110)
multiply_matrixwasn't capable of creating / allocating memory for an output image
skimage.io.imsavecaused errors with oclarrays, also in naparistandard_deviation_z_projectionhad wrong output size per default
median_boxmedian_sphereextended_depth_of_focus_variance_projectionvariance_boxfiltervariance_spherefilterz_position_of_maximum_z_projectionand aliasarg_maximum_z_projectionz_position_of_minimum_z_projectionand aliasarg_minimum_z_projectionz_position_projectionz_position_range_projection
invertwas removed from the assistant user interface. It does not do the same as in ImageJ and/or scikit-image. Thus, it might change in the future. See also #123- better documentation for
threshold_otsu,voronoi_otsu_labelingand minor others
- Relax dependencies: pyopencl-version is no longer pinned. This should simplify installation.
- Improved interoperability with cupy. You can now process images from clesperanto and cupy using syntax like
result = cle_image + cupy_image. Note: This will pull/push the image from cupy to clesperanto. Thus, it might not be very performant. - More operations have been marked as compatible with the assistant.
Added aliases for compatibility with clij 2.5
- added alias
dilate_labelsforextend_labels_with_maximum_radius - added alias
mean_intensity_mapforlabel_mean_intensity_map - added alias
mean_extension_mapforlabel_mean_extension_map - added alias
maximum_extension_mapforlabel_maximum_extension_map - added alias
extension_ratio_mapforlabel_maximum_extension_ratio_map - added alias
maximum_intensity_mapforlabel_maximum_intensity_map - added alias
minimum_intensity_mapforlabel_minimum_intensity_map - added alias
pixel_count_mapforlabel_pixel_count_map - added alias
standard_deviation_intensity_mapforlabel_standard_deviation_intensity_map
- The
touching_neighbor_count_mapdoes not count background as neighbor anymore. - Removed debug traces in
exclude_labels_with_values_within_range mean_spherereturned alwaysNone
- Standard-deviation-Z-projection is now shown in the assistant user interface
cle.operationsnow supports strings as parameters and lists of strings (as before)cle.categoriesreturns all available categories which can be used incle.operations
- the
executefunction allows providingNoneas anchor. With this, you can execute custom open.cl files located anywhere.
statistics_of_labelled_pixelshad some entries in the returned dictionary which contained misleading content. Those were removed:bbox,centroid,weighted_centroidcropreturned a 3D image with one slice when asked to crop a 2D image.
- OpenCL-buffers support now operators such as
+,-,*,/,**,+=,-=,*=,/=,**=,<,>,<=,>=,==,!=with second operand also OpenCL-buffer, numpy-array or scalar (int/float). - OpenCL-buffers now also support min, max and sum with axes 0, 1 and 2.
- Added aliases
asarrayforpushandnparrayforpull - New operation
proximal_neighbor_count_map
- Operations producing label images now automatically generate outout-images of type uint16. Operations producing binary images now automatically produce output images of type uint8. In case this breaks pre-existing workflows, consider using
cle.create_liketo create output-images of type float, which was default before this version.
sub_stackthrew an error when processing 2D images. It now warns and returns a copy.
sub_stack- activated transforms and projections in the assistant GUI
- added warnings in case operations get parameters which can cause errors, e.g. scale factor = 0
merge_touching_labelsdidn't work properly in case labels were circular connected.
apply_vector_fieldfor warping images
mean_(x/y/z)_projectionparameterdestinationis now optional
- The
linear_interpolationparameter inaffine_transform,rigid_transform,scale,translateandrotateis now functional.
- Turned off interpolation in
imshow linear_interpolation=Truecaused crashes on NVidida GPUs (#99).- If
scalewas scaling to the image center (as per default) and the output image was not the same size as the input image, parts of the image might have been removed.
This update brings many new functions, highlighing affine-transforms (rotation, translation, scaling) and filters based
on neighborhood relationships between labels. With those you can convolve measurements of cells in tissues based on their adjacency graph.
Furthermore, centroids_of_labels, statistics_of_labelled_pixels and statistics_of_background_and_labelled_pixels were GPU-accelerated.
The statistics_of_* functions, measure standard deviation of the intensity` now and results of these functions are
dictionaries which contain the same measurements as in CLIJ2.
affine_transform(yet without shearing and with nearest-neighbor interpolation only for now)centroids_of_background_and_labelsdetect_minima_boxeuclidean_distance_from_label_centroid_mapgenerate_n_nearest_neighbors_matrixgenerate_proximal_neighbors_matrixand aliasgenerate_distal_neighbors_matrixlabel_standard_deviation_intensity_maplabel_mean_extension_maplabel_maximum_extension_maplabel_maximum_extension_ratio_mapmasked_voronoi_labelingmaximum_distance_of_n_shortest_distancesmaximum_distance_of_touching_neighborsmaximum_of_n_nearest_neighbors_mapmaximum_of_proximal_neighbors_mapand aliasmaximum_of_distal_neighbors_mapmaximum_of_touching_neighbors_mapmean_of_n_nearest_neighbors_mapmean_of_proximal_neighbors_mapand aliasalias mean_of_distal_neighbors_mapmean_of_touching_neighbors_mapmerge_touching_labelsminimum_distance_of_touching_neighborsminimum_of_n_nearest_neighbors_mapminimum_of_proximal_neighbors_mapand aliasminimum_of_distal_neighbors_mapminimum_of_touching_neighbors_mapmode_of_n_nearest_neighbors_mapmode_of_proximal_neighbors_mapand aliasmode_of_distal_neighbors_mapmode_of_touching_neighbors_mappoint_index_list_to_touch_matrixread_intensities_from_mapread_intensities_from_positionsregionprops(based on scikit-image)rigid_transformrotatescalestandard_deviation_of_n_nearest_neighbors_mapstandard_deviation_of_proximal_neighbors_mapand aliasstandard_deviation_of_distal_neighbors_mapstandard_deviation_of_touching_neighbors_mapstatistics_of_imagetouch_matrix_to_adjacency_matrixtranslatevoronoi_otsu_labeling
statistics_of_labelled_pixelsandstatistics_of_background_and_labelled_pixelsproduce different output now. Instead of a scikit-image RegionProps objects, the produce a dictionary which contains lists of measurements.push_regionprops_columnalso works with the new output format. Consider updating your code to work with the dictionaries,e.g.stats['area']instead of[s.area for s in stats]. If not possible, use the new functionregionpropsretrieve results in the old formatn_closest_pointsignores the background position and the distance to labels themselves per default now. Consider passingignore_background=Falseandignore_self=Falseto go back to former functionality.
resamplehas been deprecated. Usescaleinstead.
imshowin 3D caused an errorpush_regionprops_columndidn't actually push
flag_existing_intensitieshas been renamed toflag_existing_labels. An alias to the old method was created to keep backwards-compatibility
- Operations can be categorized now. That allows generating handy user interfaces (74c82f0, d03fc2d). For example, you can retrieve a dictionary with all
background removaloperations indiced with their names by calling:
import pyclesperanto_prototype as cle
dict = cle.operations(must_have_categories=['background removal'])
- Internal support for clImages. This allows acceleated interpolation, which is necessary, e.g. for the new
resampleoperation. (075250c) - Implemented image-size independent kernel compilation. This should improve performance of operations significantly, where typically images of various sizes are processed. For example operations working with distance matrices and point lists. (d897863)
cle.functionsnow have a parameter fullargspec to enable better automatic GUI generation for them (119cee3)- We're working with a version controlled
release_notesfile now. Let's see how this evolves.
artificial_tissue_2d(08b154a)add_images(30fd15b)average_distance_of_n_closest_neighbors(e90773b)average_distance_of_n_closest_neighbors_map(4032fa7)average_distance_of_n_far_off_distances(f72e5ba)average_distance_of_n_shortest_distances(392f010)average_neighbor_distance_map(c8e939a)combine_horizontally(a527be7)combine_vertically(f8e4d70)concatenate stacks(b1b6fb0)count_touching_neighbors and create_vector_from_square_matrix(63d7865)create_from_pointlist(07cbcd2)downsample_slice_by_slice_half_median(61d6526)draw_angle_mesh_between_touching_labels(d1c5330)draw_distance_mesh_between_touch_labels(b09d717)draw_mesh_between_n_closest_neighbors(bab895f)degrees_to_radians(e73b5cb)distance_matrix_to_mesh(2f1d60e)divide_by_gaussian_background(e8f461c)exclude_labels(08e2556)exclude_labels_out_of_size_range(1fdb2f5)exclude_labels_with_values_out_of_range(62ade23)exclude_labels_with_values_within_range(2e35a4b)extend_labels_with_maximum_radius(55cf151)gamma_correction(a4f64da)generate_angle_matrix(e73b5cb)generate_binary_overlap_matrix(552369a)imread(711ebc9)imshow(711ebc9)label_centroids_to_pointlist(50ee757, 9731e3f)label_maximum_intensity_map(4bed040)label_mean_intensity_map(1de25e4)label_pixel_count_map(a8c3da1)label_spots(e7b31ef)local_minimum_touching_neighbor_count_map(9b735c1)local_maximum_touching_neighbor_count_map(9b735c1)local_mean_touching_neighbor_count_map(9b735c1)local_median/_touching_neighbor_count_map(9b735c1)local_standard_deviation_touching_neighbor_count_map(9b735c1)maximum_of_touching_neighbors(6339684)mean_of_touching_neighbors(4eae7bb)median_of_touching_neighbors(59c255a)minimum_of_touching_neighbors(86abebf)mode_of_touching_neighbors(c782f20)n_closest_points(b7fb4d7)neighbors_of_neighbors(f4ca339)pointindexlist_to_mesh(bab895f)pointlist_to_labelled_spots(7df49ee)push_regionprops(7d9b57d)radians_to_degrees(e73b5cb)resample without interpolation(4076ddd)reduce_stack(d5f32bd)spots_to_pointlist(fab24e2)standard_deviation_of_touching_neighbors(734c1b9)statistics_of_labelled_pixels(7d9b57d)statistics_of_background_and_labelled_pixels(7d9b57d)subtract_gaussian_background(cedb8e3)touching_neighbor_count_map(d92d7c4, 49fa5c9)write_values_to_positions(829bd72)
- the behaviour of
exclude_labelsis inverted now - The implementation of
label_spotschanged. It's now comparable with CLIJ2. set_nonzero_pixels_to_pixelindexwas renamed toset_non_zero_pixels_to_pixel_index
- As an answer to Issue #49,
the
pushandpullcommands warn now about deprecation. In a future release, we may remove the transposition so thatpushdoes the same aspush_zyx, andpulldoes the same aspull_zyx.
- automatic
create*functions failed when onlykwargswere passed to any operation (d033fed, 4fd6099) divide_by_gaussian_backgrounddidn't divide, it was in fact the implementation ofsubtract_gaussian_background(a6b0dd0)- before drawing a mesh, all pixels are set to 0 (9433a2f, 52572a6)
- the behaviour of
exclude_labelswas inverse (3c9f5b1, fd5d98b) - in
exclude_labelsbackground was added to label 1 in case it was kept (439dba5) detect_maxima_boxhad missing parameters compared to the CLIJ2 API