Repository navigation
v0.3.21
This release fixes several bugs that affected training results in 0.3.x and adds a proper test suite so they don't come back. We recommend everyone on 0.3.x update.
Training fixes
- Fixed the particle pixel count in 2D, which made the estimated particle fraction (pi) about 7x too high.
- Fixed negative sampling, which could draw "negative" crops on top of labeled particles.
- Fixed data loader workers producing identical batches, as found #260 by @i7878 and @bsobol.
topaz trainnow accepts TIFF and PNG micrographs, not just MRC.
Extraction and segmentation
- Fixed the 3D radius search in
topaz extract(--targetswith--dims 3). --patch-sizenow works the same way inextractandsegment: it is the size of the region scored per patch, and padding of half the model's receptive field is added automatically. Larger patches are faster but use more memory.- Fixed
segment --patch-sizecrashing, andextractcrashing when--outputis an existing directory.
Image and file handling
- MRC headers written by
preprocessanddownsamplenow have the correct size, pixel size and statistics. - PNG and JPEG images load with their original values instead of raw 0-255.
- Fixed
convert --invert-y, reading VIA.csvfiles with pandas 2, and coordinate writers droppingz_coord. - Fixed average precision for float64 inputs.
Testing and packaging
- New pytest suite (341 tests) run on Python 3.8 to 3.13 before every release. See
test/walkthrough.ipynbfor what each test checks. - Pretrained denoising models no longer need
pkg_resources. - Python 3.13 patch releases (3.13.x) can now install topaz.
Thanks to everyone who reported issues!