ucla-vast / recut Goto Github PK

Thrown sometime after writing connected_sdf.vdb

fastmarching -> fastsweeping

Some output windows will be too large to create a dense TIFF in memory especially on low RAM, or highly parallel runs with many components.

• Allow individual windows to fail and allow overall execution to continue
• Still output that components component-mask.vdb (tiny) so that the window can be generated post-hoc

The working folder has read/write permission.

However after running recut, the program aborts after 2 mins.

Anisotropic data requires upsampling the z to make it the voxel length match across all three dimensions. However upsample data that gets morphological closed in recut has different output than the same VDB library command in Houdini. The total # of somas will be far lower than actual until this is fixed

recut lazily relies on probeLeaf to do a final check whether the passed seeds correspond to an active region of topology, however if the coordinate is in a leaf whose values are all uniform it can be compressed into a tile value with results in a returned null leaf, even though the value is clearly on

Can not open in Neutube
run-14/component-10/img*.tif

Component-119

Fix: Flagged as a type B1 irregularity. Generally second point is removed (and type changed to B2), but other corrections may be necessary.

Use --preserve-topology only with --seeds X runs

The following command failed on the master branch:

time nice -n 19 recut ./mask.vdb --input-type mask --seeds /data/20220818_SW220405_05_LS_6x_1000z_stitched/soma_recut --fg-percent 0.1 --output-type eswc --close-steps 5 --open-steps 5

Thread count 256
User specified image ./mask.vdb
image: [17557, 11965, 8400]
--prune-radius was set to: 1 by calculating the anisotropic factor of --voxel-size
All outputs will be written to: ./run-2
Only 41 of 41 seeds in directory have an active voxel and are connected to a component in the provided image
recut: /build/r7qczr56i0ida94i5gl10alam5mlv8gk-source/src/recut.hpp:3249: void Recut<image_t>::operator()() [with image_t = short unsigned int]: Assertion `((void)"Root pairs must be set before beginning reconstruction", !seeds.empty())' failed.
Aborted

Using recut starting from 2D tiff series on a whole brain image consumes 700GB more RAM compared with the time starting with a mask.vdb file.

tested commands:

recut /path/to/2D_tiff_series --fg-percent 0.1 --output-type eswc --open-steps 6 --close-steps 6

recut /path/to/mask.vdb --input-type mask --fg-percent 0.1 --output-type eswc --open-steps 6 --close-steps 6

Do so immediately such that output windows have as much memory ceiling as possible

Runtime error on large scale reconstruction "parent is also invalid". Occurs in a tiny fraction of components but will stop execution entirely during SWC generation.

Can be fixed quickly by convert directly to graph, skipping the mesh manifold conversion which is largely for visualization/observability

tested command:
time nice -n 19 recut /data/20220725_17_41_55_SW220510_02_LS_6x_1000z_stitched/Ex_488_Em_525_6x_8bit2bsh_tif_deconvolved_flip_up --seeds /data/20220725_17_41_55_SW220510_02_LS_6x_1000z_stitched/soma_recut --fg-percent 0.1 --output-type eswc --open-steps 4 --close-steps 1

Error:
Start merge Grid finalize time: 102.57 s Finished stage: convert Finished total updating within 3564.17 sec recut: /build/521gxvnvw53pza2mcdkl8b5300pz0s5j-source/src/recut.hpp:3255: void Recut<image_t>::operator()() [with image_t = short unsigned int]: Assertion ((void)"Root pairs must be set before beginning reconstruction", !root_pairs.empty())' failed. Aborted

Nodes with radius below 1 get written into SWC as 0.xxx

That combines the neighbor clearing strategy of force, but that uses the actual computed radius size

Small side branch not long enough to "escape" the current diameter of parent segment.
Fix: Type A irregularity; Point is automatically removed.

Runtime error on large scale reconstruction "passed radii value of 0 is invalid". Occurs in a tiny fraction of components but will stop outputting more SWCS entirely instead of logging and continuing

• Keep temporarily disabled for output windows to check serial memory reqs
• Requires forcing GEL to execute single threaded

--image-offsets coupled with --image-lengths allows you to crop an image to a specific bounding box, for example a bounding box containing only the proofread seeds or connected_sdf from a previous run. This could boost the signal to noise ration due to high saturation in regions outside of the bounding box.

Bug is when passing offsets of x and y, currently only z works.