In train_data_config.json, a mask for each even/odd pair can now be specified, which will limit the coordinate extraction to regions within the mask.
Example:
5 K2 tomograms from EVN/ODD MotionCor2 frames aligned using AreTomo and reconstructed at bin 4 (5.49 A/pix) using imod tilt (tomotools workflow). Dimensions: 959 x 927 x 1093.
Boundary models created in imod slicer view, converted to binary masks using mod2mask.
Binary mask covers only ~25% of total tomogram value
Extraction, training and prediction with default parameters in cryoCARE v0.3 - either wt or patched with masks
Result: In the xy plane, differences are minimal between either strategy (masked or unmasked), probably because 1200 subtomograms per tomogram at 72 px side length anyways already covers 45% of total voxels. In the xz plane, reconstruction artifacts are much reduced after masked training, so probably there is some advantage:
Checks:
Coordinates for extraction do respect the mask (as extracted from train_data.npz/coords.npy), see histogram of z positions below (masked in blue, unmasked in orange).
Extraction with patched cryoCARE_extract_train_data.py also works if mask is not specified in train_data_config.json.
In
train_data_config.json, a mask for each even/odd pair can now be specified, which will limit the coordinate extraction to regions within the mask.Example:
MotionCor2frames aligned usingAreTomoand reconstructed at bin 4 (5.49 A/pix) using imodtilt(tomotools workflow). Dimensions: 959 x 927 x 1093.imodslicer view, converted to binary masks using mod2mask.cryoCAREv0.3 - eitherwtor patched withmasksChecks:
cryoCARE_extract_train_data.pyalso works ifmaskis not specified intrain_data_config.json.