Repository navigation
Sources.py: quick wins, split into a package, and API unification #150
Description
src/machinevisiontoolbox/Sources.py is about 4250 lines and holds every image source class: VideoFile, VideoCamera, WebCam, EarthView, FileCollection, FileArchive, ImageSequence, TensorStack, PointCloudSequence, LabelMeReader, and the ROS classes (ROSTopic, SyncROSStreams, ROSBag). The ROS code alone is roughly 40% of the file, and depends one-way on the rest. This issue collects the work to tidy it up. Contributions are welcome: items are ordered, and the first group needs no design decisions.
The RealSense ROSBag decoding bug is tracked separately (#151), because it is a user-facing defect and can be fixed independently.
1. Quick wins (no API change, one small PR each is fine)
-
VideoFile: raise a clear error when the file cannot be opened, instead of failing later withAttributeErroron aNonecapture. -
LabelMeReader: raise on bad input instead ofreturn None. -
ImageSource: thehasattr(self, "__len__")checks (in__init__and the iteration setup) are dead code, since the class defines the attribute; replace them with a real test. - Add timeouts to
urllib.request.urlretrieve/urlopen(ROSBagURL download,LabelMeReader) and to theEvent.wait()calls inROSTopic, so a stalled connection cannot hang forever. - Delete the unused
_PC2_DTYPEtable and theif __name__ == "__main__":block at the bottom. - Fix docstring errors:
prop/propertyin the camera-property methods, the duplicatedshow_allowedentry inROSBag.print,tol/tolerance, and the "coolumn" typo. -
docs/source/sources.rstdocuments aread()method that does not exist;docs/source/ros.rstcall signatures need checking against the code. - Collapse the repeated
colororderbranching (about ten copies) into one helper that wraps an array as anImage. -
ROSBag: replaceprint()withlogging(about 30printcalls in the file).
2. Split into a package (pure move, its own PR, no behaviour change)
git mv Sources.py Sources/__init__.py, then move classes into _base.py (ImageSource and shared helpers), video.py, files.py, archive.py, web.py, tensor.py, labelme.py, pointcloud.py and ros.py.
-
__init__.pymust re-export every public name and the private availability flags that tests read viagetattr. - Update test patch targets that name
Sources.safe_plt_show, and check the Sphinxcurrentmoduledirectives. - Verify with an import of every public name from both
machinevisiontoolboxandmachinevisiontoolbox.Sources, thentests/test_image_sources.pyandtests/test_ros.py(needs therosextra). - Optional follow-up: lazy-import
rosbags/roslibpysoimport machinevisiontoolboxstops paying for them.
Please do the split after the quick wins, in a PR containing nothing else, so it is easy to review as a pure move. Other work (for example making optional data files downloadable on demand) will also edit this file's call sites, so landing the split first avoids painful rebases.
3. API unification (changes behaviour, needs design discussion first)
- Shared
loop,rgb/colororder,dtypeand start/stop/step options on the base class. -
__iter__returns a fresh iterator each time. - One live-source API:
close()plus context-manager support everywhere, retiregrab(). - Lock
ROSTopic's latest-frame and timestamp state.
Any of these should come with deprecation warnings for the old behaviour. Please open a discussion or comment here before starting.
4. After the split
- Annotate the split modules (the parameter annotations for
Sources.pyare tracked in Type-hint coverage: ~89% overall, remaining gaps in ImageCore, Sources, blocks, bin #60; about 50 are missing today). - Add mocked tests for
VideoCamera,WebCamandEarthView.
Notes for contributors
- Branch from
mainand open a PR; titles follow Conventional Commits (fix:,refactor:,docs:). - Please keep each PR to one of the items above.
- Related open work: EarthView: read key from GOOGLE_MAPS_API_KEY to avoid clash with Gemini's GOOGLE_API_KEY #125 is fixed by PR fix: EarthView reads GOOGLE_MAPS_API_KEY first #149 (
EarthViewkey lookup), which touches this file; it should land first.
Activity
fix: report video file initialisation failures
fix: Sources.py: quick wins, split into a package, and API unification
fix: LabelMeReader raises on unreadable input instead of returning None