Closed uzaymacar closed 2 years ago
The first problem detected with the config files is that the center crop, length, and stride parameters are not suitable for the spine-generic-processed
dataset. Therefore, below is a quick exploration of the resolutions and sizes of the image data across the three contrasts T1w, T2*, and T2w.
RESOLUTIONS: Counter({(1.0, 1.0, 1.0): 254})
SIZES: Counter({(192, 260, 320): 169, (192, 320, 320): 49, (188, 320, 320): 33, (190, 260, 320): 1, (192, 170, 320): 1, (192, 254, 313): 1})
RESOLUTIONS: Counter({(0.5, 0.5, 5.0): 243, (0.4, 0.4, 5.0): 8, (0.5, 0.5, 7.0): 1, (0.5, 0.5, 6.0): 1, (0.5, 0.5, 6.5): 1})
SIZES: Counter({(448, 448, 15): 195, (320, 320, 15): 44, (512, 512, 15): 8, (448, 448, 24): 7})
RESOLUTIONS: Counter({(0.8, 0.8, 0.8): 254})
SIZES: Counter({(64, 320, 320): 172, (63, 320, 320): 49, (56, 320, 320): 33})
Latest commit e8aef0a introduces the following training hyperparameter updates across all configs:
super_duper
appended at beginning to differentiate from #18 slice_filter_params
with patch_filter_params
-- more to come about this, follow https://github.com/ivadomed/ivadomed/issues/1163.(224, 256, 320)
for nowTo briefly justify the last point, we are inputting the whole volume during both training and inference to the model at each pass similar to how we did with SC segmentation in model_seg_ms_mp2rage project. The value (224, 256, 320)
was chosen by resampling all volumes across all contrasts to 1x1x1 mm^3 resolution and getting the maximum size across each dimension x, y, and z. This value shouldn't create a problem in inference in terms of RAM overflow as discussed in model_seg_ms_mp2rage/issues/29.
This is probably not the most efficient nor ideal solution. However it is chosen initially as the ideas we have discussed in the last meeting (e.g., variable center-crop and cascaded architecture; see meeting minutes) will take longer to implement as it requires modifications to ivadomed
codebase. I will be working on the required modifications in parallel while we get baselines with the described approach.
This PR adds 4 config files for training super-duper methods, which refers to using a single label for all contrasts that is sampled in the native space of the image:
seg_sc_all.json
which is trained on T1w, T2*, and T2w contrastsseg_sc_t1w.json
which is trained on T1w contrastseg_sc_t2star
which is trained on T2* contrastseg_sc_t2w
which is trained on T2w contrastAll config files in this PR use the
spine-generic-processed
dataset (version:4d60963478566a472897114d0c53c01cb327c48f
) and thelabels_softseg
derivative folder.