v0.3.0.summary.md

0.3.0 - Summary Of Changes

Improved Segmentation Map Augmentation (#302)

The segmentation map augmentation was previously previously a wrapper around heatmap augmentation. This patch introduces independent methods for segmentation map augmentation. This makes the augmentation of such inputs faster and more memory efficient. The internal representation (int instead of floats) also becomes more intuitive.

This improvement leads to some breaking changes. To adapt to the new version, the following steps should be sufficient for most users:

• Rename all calls of SegmentationMapOnImage to SegmentationMapsOnImage (Map -> Maps).
• Rename all calls of SegmentationMapsOnImage.get_arr_int() to SegmentationMapsOnImage.get_arr().
• Remove the argument nb_classes from all calls of SegmentationMapsOnImage.
• Remove the argument background_threshold from all calls as it is no longer supported.
• Remove the argument draw_foreground_mask from all calls of SegmentationMapsOnImage.draw_on_image() as it is no longer supported.
• Ensure that the input array to SegmentationMapsOnImage is always an int-like (int, uint or bool). Float arrays are now deprecated.
• Adapt all calls SegmentationMapsOnImage.draw() and SegmentationMapsOnImage.draw_on_image(), as both of these now return a list of drawn images instead of a single array. (For a segmentation map array of shape (H,W,C) they return C drawn images. In most cases C=1, so simply call draw()[0] or draw_on_image()[0].)
• Ensure that if SegmentationMapsOnImage.arr is accessed anywhere, the respective code can handle the new int32 (H,W,#maps) array form. Previously, it was float32 and the channel-axis had the same size as the max class id (+1) that could appear in the map.
• Ensure that calls of <augmenter>.augment() or <augmenter>() that provide segmentation maps as numpy arrays (i.e. bypassing SegmentationMapsOnImage) use the shape (N,H,W,#maps) as (N,H,W) is no longer supported.

New RNG System (#375, #408)

numpy 1.17 introduces a new API for random number generation. This patch adapts imgaug to automatically use the new API if it is available and fall back to the old one otherwise. To achieve that, the module imgaug.random is introduced, containing the new standard random number generator imgaug.random.RNG. You can create a new RNG using a seed value via RNG(seed) and it will take care of the rest. It supports all sampling functions that numpy.random.RandomState and numpy.random.Generator support. This new random number generator is now supposed to be used wherever previously numpy.random.RandomState would have been used. (For most users, this shouldn't change anything. Integer seeds are still supported. If you used RandomState anywhere, that is also still supported.)

Breaking changes related to this patch:

• imgaug now uses a different seed at each run of the library. Previously, a fixed seed was used for each run, leading to the same agumentations. That confused some users as it differed from numpy's behaviour. The new "dynamic" seed is derived from numpy's seed and hence seeding numpy will also lead to imgaug being seeded. (It is not recommended to rely on that behaviour as it might be changed in the future. Use imgaug.random.seed() to set a custom seed.)
• The constants imgaug.SEED_MIN_VALUE and imgaug.SEED_MAX_VALUE were removed. They are now in imgaug.random.
• The constant imgaug.CURRENT_RANDOM_STATE was removed. Use imgaug.random.get_global_rng() instead.

Other Changes Related to numpy 1.17 (#302)

numpy 1.17 uses a new implementation of clip(), which turns int64 values into float64 values. As a result, it is no longer safe to use int64 in many augmenters and other functions/methods and hence these inputs are now rejected. This affects at least ReplaceElementwise and thereby Dropout, CoarseDropout, Salt, Pepper, SaltAndPepper, CoarseSalt, CoarsePepper and CoarseSaltAndPepper. See the ReadTheDocs documentation page about dtype support for more details.

In relation to this change, parameters in imgaug.parameters that previously returned int64 were modified to now return int32 instead. Analogously, float64 results were changed to float32.

New Augmenters

The following new augmenters were added to the library:

Canny edge detection (#316):

• imgaug.augmenters.edges.Canny. Performs canny edge detection and colorizes the resulting binary image in random ways.

Pooling (#317):

• imgaug.augmenters.edges.AveragePooling. Performs average pooling using a given kernel size. Very similar to AverageBlur.
• imgaug.augmenters.edges.MaxPooling. Performs maximum pooling using a given kernel size.
• imgaug.augmenters.edges.MinPooling. Analogous.
• imgaug.augmenters.edges.MedianPooling. Analogous.

Hue and Saturation (#210, #319):

• imgaug.augmenters.color.WithHueAndSaturation. Apply child augmenters to images in HSV colorspace. Automatically accounts for the hue being in angular representation.
• imgaug.augmenters.color.AddToHue. Adds a defined value to the hue of each pixel in input images.
• imgaug.augmenters.color.AddToSaturation. Adds a defined value to the saturation of each pixel in input images.
• imgaug.augmenters.color.MultiplyHueAndSaturation. Multiplies the hue and/or saturation of all pixels in input images.
• imgaug.augmenters.color.MultiplyHue. Analogous, affects always only the hue.
• imgaug.augmenters.color.MultiplySaturation. Analogous, affects always only the saturation.

Color Quantization (#347):

• imgaug.augmenters.color.UniformColorQuantization. Uniformly splits all possible colors into N different ones, then finds for each pixel in an image among the N colors the most similar one and replaces that pixel's color with the quantized color.
• imgaug.augmenters.color.KMeansColorQuantization. Groups all colors in an each into N different ones using k-Means clustering. Then replaces each pixel'S color, analogously to UniformColorQuantization.

Voronoi (#348):

• imgaug.augmenters.segmentation.Voronoi. Queries a point sampler to generate a large number of (x,y) coordinates on an image. Each such coordinate becomes a voronoi cell. All pixels within the voronoi cell are replaced by their average color. (Similar to Superpixels, this augmenter also supports to only replace p% of all cells with their average color.)
• imgaug.augmenters.segmentation.UniformVoronoi. Shortcut to call Voronoi with a uniform points sampler. That sampler places N points on an image using uniform distributions (i.e. they are randomly spread over the image.)
• imgaug.augmenters.segmentation.RegularGridVoronoi. Shortcut to call Voronoi with a regular grid points sampler. That points sampler generates coordinate on a regular grid with H rows and W cols. Some of these points can be randomly dropped to generate a less regular pattern.
• imgaug.augmenters.segmentation.RelativeRegularGridVoronoi. Same as RegularGridVoronoi, but instead of using absolute numbers for H and W, they are defined as relative amounts w.r.t. image shapes, leading to more rows/cols on larger images.

New Augmentation Functions

One of the long term goals of the library is to move as much augmentation logic as possible out of Augmenter instances and into functions. This patch therefore adds several new augmentation functions:

• imgaug.min_pool(). #369
• imgaug.median_pool(). #369
• augmenters.segmentation.segment_voronoi(). #348
• augmenters.flip.fliplr(). #385
• augmenters.flip.flipud(). #385
• augmenters.color.change_colorspace_(). #409
• augmenters.color.change_colorspace_batch_(). #409
• augmenters.arithmetic.add_scalar(). #411
• augmenters.arithmetic.add_elementwise(). #411
• augmenters.arithmetic.replace_elementwise_(). #411
• augmenters.arithmetic.compress_jpg(). #411

Colorspace Changes (#409)

The color space naming within the library had become rather messy in the past as there were many colorspace-related augmenters, with some of them not using constants for colorspace names/IDs and others defining their own ones. This patch introduces a unified colorspace naming system for which the following constants were added:

• imgaug.CSPACE_RGB
• imgaug.CSPACE_BGR
• imgaug.CSPACE_GRAY
• imgaug.CSPACE_CIE
• imgaug.CSPACE_YCrCb
• imgaug.CSPACE_HSV
• imgaug.CSPACE_HLS
• imgaug.CSPACE_Lab
• imgaug.CSPACE_Luv
• imgaug.CSPACE_YUV
• imgaug.CSPACE_ALL

All colorspace-related augmenters should now support these constants.

Additionally, support for rarely used colorspaces -- mainly CIE, YCrCb, Luv and YUV -- was previously unverified or non-existent. These colorspaces are now tested for the underlying transformation functions and should be supported by most colorspace-related augmenters. (Some augmenters may still define their own subset of actually sensible colorspaces and only accept these.)

Setting limits on memory usage of background augmentation (#305, #417)

The methods imap_batches() and imap_batches_unordered() of imgaug.multicore.Pool have now the new argument output_buffer_size. The argument set the maximum number of batches that may be handled anywhere in the augmentation pipeline at a given time (i.e. in the steps "loaded and waiting", "in augmentation" or "augmented and waiting"). It denotes the total number of batches over all processes. Setting this argument to an integer value avoids situations where Pool eats up all the available memory due to the data loading and augmentation running faster than the training.

Augmenter.augment_batches() now uses a default value of 10*C for output_buffer_size, where C is the number of available logical CPU cores.

Performance Related Changes

The algorithms for Fliplr and Flipud were reworked to be as fast as possible. In practice this should have no noticeable effects as both augmenters were already very fast. (#385)

Furthermore, all assert statements within the library were changed from do_assert() to standard assert statements. This is a bit less secure (as assert statements can be optimized away), but should have a small positive impact on the performance. (#387)

Large parts of the library were also refactored to reduce code duplication and decrease the complexity of many functions. This should make future improvements easier, but is expected to have a very small negative impact on the performance due to an increased number of function calls. It is also expected that numpy 1.17 can make some operations slower. This is because (a) creating and copying random number generaters has become slower and (b) clip() overall seems to be slower.

Improved Error Messages (#366, #367, #387)

imgaug uses quite many assert statements and other checks on input data to fail early instead of late. This is supposed to improve usability, but that goal was not always reached as many errors had no associated error messages. This patch changes that. Now, all assert statements and other checks have an associated error message. This should protect users from having to wade through the library's code in order to understand the root cause of errors.

(Almost) All Augmenters Are Now Classes (#396)

Some augmenters were previously defined as functions returning other augmenters with appropriate settings. This could lead to confusing effects, where seemingly instantiating an augmenters would lead to the instantiation of a completely different augmenter. Hence, most of these augmenters were switched from functions to classes. (The classes are now inheriting from the previously returned augmenters, i.e. instanceof checks should still work.) This affects: AdditiveGaussianNoise, AdditiveLaplaceNoise, AdditivePoissonNoise, Dropout, CoarseDropout, ImpulseNoise, SaltAndPepper, CoarseSaltAndPepper, Salt, CoarseSalt, Pepper, CoarsePepper, SimplexNoiseAlpha, FrequencyNoiseAlpha, MotionBlur, MultiplyHueAndSaturation, MultiplyHue, MultiplySaturation, AddToHue, AddToSaturation, Grayscale, GammaContrast, SigmoidContrast, LogContrast, LinearContrast, Sharpen, Emboss, EdgeDetect, DirectedEdgeDetect, OneOf, AssertLambda, AssertShape, Pad, Crop, Clouds, Fog and Snowflakes.

Not yet switched are: InColorspace (deprecated), ContrastNormalization (deprecated), HorizontalFlip (pure alias for Fliplr), VerticalFlip (pure alias for Flipud) and Scale (deprecated).

Augmenters are now more robust towards unusual axis-sizes (#428, #433)

Feeding images with height and/or width of 0 or a channel axis of size 0 into augmenters would previously often result in crashes. This was also the case for input arrays with more than 512 channels. Some of these errors also included segmentation faults or endlessly hanging programs. Most augmenters and helper functions were modified to be more robust towards such unusual inputs and will no longer crash.

It is still good practice to avoid such inputs. Note e.g. that some helper functions -- like drawing routines -- may still crash. The unittests corresponding to this change also only cover image data. Using other inputs, e.g. segmentation maps, might still induce problems.

Other New Functions

The following (public) functions were added to the library (not listing functions that were already mentioned above):

• Added imgaug.is_np_scalar(). #366
• Added dtypes.normalize_dtypes(). #366
• Added dtypes.normalize_dtype(). #366
• Added dtypes.change_dtypes_(). #366
• Added dtypes.change_dtype_(). #366
• Added dtypes.increase_itemsize_of_dtype(). #366
• Added imgaug.warn() function. #367
• Added imgaug.compute_paddings_to_reach_multiples_of(). #369
• Added imgaug.pad_to_multiples_of(). #369
• Added augmentables.utils.copy_augmentables. #410
• Added validation.convert_iterable_to_string_of_types(). #413
• Added validation.is_iterable_of(). #413
• Added validation.assert_is_iterable_of(). #413
• Added random.supports_new_rng_style(). #375
• Added random.get_global_rng(). #375
• Added random.seed(). #375
• Added random.normalize_generator(). #375
• Added random.normalize_generator_(). #375
• Added random.convert_seed_to_generator(). #375
• Added random.convert_seed_sequence_to_generator(). #375
• Added random.create_pseudo_random_generator_(). #375
• Added random.create_fully_random_generator(). #375
• Added random.generate_seed_(). #375
• Added random.generate_seeds_(). #375
• Added random.copy_generator(). #375
• Added random.copy_generator_unless_global_generator(). #375
• Added random.reset_generator_cache_(). #375
• Added random.derive_generator_(). #375
• Added random.derive_generators_(). #375
• Added random.get_generator_state(). #375
• Added random.set_generator_state_(). #375
• Added random.is_generator_equal_to(). #375
• Added random.advance_generator_(). #375
• Added random.polyfill_integers(). #375
• Added random.polyfill_random(). #375

Other New Classes and Interfaces

The following (public) classes were added (not listing classes that were already mentioned above):

• Added augmenters.edges.IBinaryImageColorizer. #316
• Added augmenters.edges.RandomColorsBinaryImageColorizer. #316
• Added augmenters.segmentation.IPointsSampler. #348
• Added augmenters.segmentation.RegularGridPointsSampler. #348
• Added augmenters.segmentation.RelativeRegularGridPointsSampler. #348
• Added augmenters.segmentation.DropoutPointsSampler. #348
• Added augmenters.segmentation.UniformPointsSampler. #348
• Added augmenters.segmentation.SubsamplingPointsSampler. #348
• Added testutils.ArgCopyingMagicMock. #413

The image colorization is used for Canny to turn binary images into color images. The points samplers are currently used within Voronoi.

Refactorings

Due to fast growth of the library in the past, a significant amount of messy code had accumulated. To fix that, a lot of time was spend to refactor the code throughout the whole library to reduce code duplication and improve the general quality. This also included a rewrite of many outdated docstrings. There is still quite some mess remaining, but the current state should make it somewhat easier to add future improvements.

As part of the refactorings, a few humongously large unittests were also split up into many smaller tests. The library has now around 3000 unique unittests (i.e. each unittest function is counted once, even it is called many times with different parameters).

Related PRs:

• #302, #319, #328, #329, #330, #331, #332, #333, #334, #335, #336, #351, #352, #353, #354, #355, #356, #359, #362, #366, #367, #368, #369, #389, #397, #401, #402, #403, #407, #409, #410, #411, #413, #419

Deprecated

The following functions/classes/arguments are now deprecated:

• Function imgaug.augmenters.meta.clip_augmented_image_. Use imgaug.dtypes.clip_() or numpy.clip() instead. #398
• Function imgaug.augmenters.meta.clip_augmented_image. Use imgaug.dtypes.clip_() or numpy.clip() instead. #398
• Function imgaug.augmenters.meta.clip_augmented_images_. Use imgaug.dtypes.clip_() or numpy.clip() instead. #398
• Function imgaug.augmenters.meta.clip_augmented_images. Use imgaug.dtypes.clip_() or numpy.clip() instead. #398
• Function imgaug.normalize_random_state. Use imgaug.random.normalize_generator instead. #375
• Function imgaug.current_random_state. Use imgaug.random.get_global_rng instead. #375
• Function imgaug.new_random_state. Use class imgaug.random.RNG instead. #375
• Function imgaug.dummy_random_state. Use imgaug.random.RNG(1) instead. #375
• Function imgaug.copy_random_state. Use imgaug.random.copy_generator instead.
• Function imgaug.derive_random_state. Use imgaug.random.derive_generator_ instead. #375
• Function imgaug.normalize_random_states. Use imgaug.random.derive_generators_ instead. #375
• Function imgaug.forward_random_state. Use imgaug.random.advance_generator_ instead. #375
• Augmenter imgaug.augmenters.arithmetic.ContrastNormalization. Use imgaug.augmenters.contrast.LinearContrast instead. #396
• Argument X in imgaug.augmentables.kps.compute_geometric_median(). Use argument points instead. #402
• Argument cval in imgaug.pool(), imgaug.avg_pool() and imgaug.max_pool(). Use pad_cval instead. #369

Dependencies

The following changes were made to the dependencies of the library:

• Increased minimum version requirement for scikit-image to 0.14.2. #377, #399
• Changed dependency opencv-python to opencv-python-headless. This should improve support for some system without GUIs. #324
• Added dependency pytest-subtests for the library's unittests. #366

conda-forge

The library was added to conda-forge so that it can now be installed via conda install imgaug. (The conda-forge channel must be added first, see installation docs or README.) #320 #339

Fixes

• Fixed an issue with Polygon.clip_out_of_image(), which would lead to exceptions if a polygon had overlap with an image, but not a single one of its points was inside that image plane.
• Fixed multicore methods falsely not accepting augmentables.batches.UnnormalizedBatch.
• Rot90 now uses subpixel-based coordinate remapping. I.e. any coordinate (x, y) will be mapped to (H-y, x) for a rotation by 90deg. Previously, an integer-based remapping to (H-y-1, x) was used. Coordinates are e.g. used by keypoints, bounding boxes or polygons.
• augmenters.arithmetic.Invert
  • [rarely breaking] If min_value and/or max_value arguments were set, uint64 is no longer a valid input array dtype for Invert. This is due to a conversion to float64 resulting in loss of resolution.
  • Fixed Invert in rare cases restoring dtypes improperly.
• Fixed dtypes.gate_dtypes() crashing if the input was one or more numpy scalars instead of numpy arrays or dtypes.
• Fixed augmenters.geometric.PerspectiveTransform producing invalid polygons (more often with higher scale values). #338
• Fixed errors caused by external/poly_point_isect_py2py3.py related to floating point inaccuracies (changed an epsilon from 1e-10 to 1e-4, rounded some floats). #338
• Fixed Superpixels breaking when a sampled n_segments was <=0. n_segments is now treated as 1 in these cases.
• Fixed ReplaceElementwise both allowing and disallowing dtype int64. #346
• Fixed BoundingBox.deepcopy() creating only shallow copies of labels. #356
• Fixed dtypes.change_dtypes_() #366
  • Fixed argument round being ignored if input images were a list.
  • Fixed failure if input images were a list and dtypes a single numpy dtype function.
• Fixed dtypes.get_minimal_dtype() failing if argument arrays contained not exactly two items. #366
• Fixed calls of CloudLayer.get_parameters() resulting in errors. #309
• Fixed SimplexNoiseAlpha and FrequencyNoiseAlpha not handling sigmoid argument correctly. #343
• Fixed SnowflakesLayer crashing for grayscale images. #345
• Fixed Affine heatmap augmentation crashing for arrays with more than four channels and order!=0. #381
• Fixed an outdated error message in Affine. #381
• Fixed Polygon.clip_out_of_image() crashing if the intersection between polygon and image plane was an edge or point. #382
• Fixed Polygon.clip_out_of_image() potentially failing for polygons containing two or fewer points. #382
• Fixed Polygon.is_out_of_image() returning wrong values if the image plane was fully contained inside the polygon with no intersection between the image plane and the polygon edge. #382
• Fixed Fliplr and Flipud using for coordinate-based inputs and image-like inputs slightly different conditions for when to actually apply augmentations. #385
• Fixed Convolve using an overly restrictive check when validating inputs for matrix w.r.t. whether they are callables. The check should now also support class methods (and possibly various other callables). #407
• Fixed CropAndPad, Pad and PadToFixedSize still clipping cval samples to the uint8. They now clip to the input array's dtype's value range. #407
• Fixed WithColorspace not propagating polygons to child augmenters. #409
• Fixed WithHueAndSaturation not propagating segmentation maps and polygons to child augmenters. #409
• Fixed AlphaElementwise to blend coordinates (for keypoints, polygons, line strings) on a point-by-point basis following the image's average alpha value in the sampled alpha mask of the point's coordinate. Previously, the average over the whole mask was used and then either all points of the first branch or all of the second branch were used as the augmentation output. This also affects SimplexNoiseAlpha and FrequencyNoiseAlpha. #410
• Fixed many augmenters and helper functions producing errors if the height, width and/or channels of input arrays were exactly 0 or the channels were >512. See further above for more details. #433
• Fixed Rot90 not supporting imgaug.ALL. #434
• Fixed PiecewiseAffine possibly generating samples for non-image data when using absolute_scale=True that were not well aligned with the corresponding images. #437