Python library for converting between deep-learning annotation formats.
wai.annotations core module, containing core data structures and basic data loading and preprocessing techniques.
The following sources are available:
from-audio-files-ac: dummy reader that turns audio files into a classification dataset.
from-audio-files-sp: dummy reader that turns audio files into a speech dataset.
from-images-ic: dummy reader that turns images into an image classification dataset.
from-images-is: dummy reader that turns images into an image segmentation dataset.
from-images-od: dummy reader that turns images into an object detection dataset.
The following inline stream processors (ISPs) area available:
check-duplicate-filenames: causes the conversion stream to halt when multiple dataset items have the same filename
clean-transcript: ISP that cleans speech transcripts.
coerce-box: converts all annotation bounds into box regions
coerce-mask: converts all annotation bounds into polygon regions
convert-image-format: converts images from one format to another
dimension-discarder: removes annotations which fall outside certain size constraints
discard-invalid-images: discards images that cannot be loaded (e.g., corrupt image file or annotations with no image)
discard-negatives: discards negative examples (those without annotations) from the stream
filter-labels: filters detected objects down to those with specified labels.
filter-metadata: filters detected objects based on their meta-data.
label-present: keeps or discards images depending on whether annotations with certain label(s) are present. Checks can be further tightened by defining regions in the image that annotations must overlap with (or not overlap at all).
map-labels: maps object-detection labels from one set to another
passthrough: dummy ISP which has no effect on the conversion stream
polygon-discarder: removes annotations with polygons which fall outside certain point limit constraints
remove-classes: removes classes from classification/image-segmentation instances
rename: ISP that renames files.
sample: ISP that selects a subset from the stream.
strip-annotations: ISP which removes annotations from instances
The following cross-domain converters (XDCs) are available: * od-to-ic: converts image object-detection instances into image classification instances * od-to-is: converts image object-detection instances into image segmentation instances
The following sinks are available:
to-audio-files-ac: dummy writer that just outputs audio files from classification datasets.
to-audio-fileS-sp: dummy writer that just outputs audio files from speech datasets.
to-images-ic: dummy writer that just outputs images from image classification datasets.
to-images-is: dummy writer that just outputs images from image segmentation datasets.
to-images-od: dummy writer that just outputs images from object detection datasets.
to-void-ac: consumes audio classification instances without writing them.
to-void-ic: consumes image classification instances without writing them.
to-void-is: consumes image segmentation instances without writing them.
to-void-od: consumes instances without writing them.
to-void-sp: consumes instances without writing them.
Updated wai.common requirement to v0.0.41.
Added spectrum classification domain.
Added WriteLabels ISP which can read labels from classification, object-detection and image-segmentation streams, and writes them to disk.
FilterLabels ISP now treats elements as negative ones if no labels left after filtering (in order to use discard-negatives in pipeline); also works on image classification domain now as well
FilterLabels ISP can filter out located objects that don’t fall within a certain region (x,y,w,h - normalized or absolute) using a supplied IoU threshold; useful when concentrating on annotations in the center of an image, e.g., for images generated with the subimages ISP (object detection domain only)
logging._LoggingEnabled module now sets the numba logging level to WARNING
logging._LoggingEnabled module now sets the shapely logging level to WARNING
core.domain.Data class now stores the path of the file as well
Rename ISP allows renaming of files, e.g., for disambiguating across batches
batch_split.Splitter now handles cases when the regexp does not produce any matches (and outputs a warning when in verbose mode)
Added LabelPresent ISP, which skips object detection images that do not have specified labels (or if annotations do not overlap with defined regions; can be inverted).
Using wai.common==0.0.40 now to avoid parse error output when accessing poly_x/poly_y meta-data in LocatedObject instances when containing empty strings.
The CleanTranscript ISP can be used to clean up speech transcripts.
Bug fix for splitting where split-scheduling was calculated with swapped iteration order, leading to runs of splits rather than desired interleaving. Added –no-interleave flag to re-enable bug for backwards compatibility.
Added new audio domain for classification using suffix -ac
Added dataset reader for audio files: from-audio-files-sp, from-audio-files-ac
Added dataset writer for audio files: to-audio-files-sp, to-audio-files-ac
Added dummy sink for audio files: to-void-ac
Added ISP for selecting a sub-sample from the stream: sample
Added discard-invalid-images ISP for removing corrupt images or annotations with no image attached.
Added batch-split sub-command for splitting individual batches of annotations into subsets like train/test/val. Supports grouping of files within batches (eg multiple images of the same object).
Added filter-metadata ISP for filtering object detection.
Restricted maximum characters per line in help output to 100 to avoid long help strings to become unreadable.
The polygon-discarder now annotations that either have no polygon or invalid polygons.
Added descriptions to the help screens of the main commands.
The ImageSegmentationAnnotation class now outputs the unique values in its exception when there are more unique values than labels
The Data class (module: wai.annotations.core.domain) now outputs a warning message if a file cannot be read; also added LoggingEnabled mixin.
Image segmentation annotations received new label_images property that returns images per label (as a dictionary); moved from the ToLayerSegments conversion (wai.annotations.layersegments)
Added dataset readers that generate dummy datasets from images: from-images-ic, from-images-is, from-images-od
Added dataset writers that just output the images from datasets: to-images-ic, to-images-is, to-images-od
Added ISP for discarding polygons that either have too few or too many points (polygon-discarder)
Added –verbose flag to dimension-discarder ISP for outputting information when an annotation gets discarded.
Added cross-domain converter which reduces object-detection datasets to image classification by looking for single/majority objects/labels.
Added void output formats for each domain which discard the conversion results.
Added option to LocalFilenameSource which tells it to write the files it reads to a list-file.
Initial release after separation from wai.annotations main repo.
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Hashes for wai.annotations.core-0.2.2.tar.gz