FrancescoCasalegno
commented
2 years ago Update
This issues seems to be related to the fact that different feature extraction methods produce a different number of output cells for the same dataset. That is, feature extraction probably fails for some morphologies.
After fixing some bugs (see in particular #73), we now have only two datasets where we can see different number of cells. Indeed, running (on feed8f456499b0cc1c234e000054a294d23668ef) the following snippet (run inside dvc/extract-features)
print("dataset n_cells")
print("----------------------------------------------")
for dataset in sorted(Path.cwd().glob("*")):
n_cells = None
ss = set()
for dendrite in sorted(dataset.glob("*")):
for method in sorted(dendrite.glob("*")):
if method.is_file():
continue
n_cells_new = len(sorted(method.glob("*")))
ss.add(n_cells_new)
ss = ", ".join(str(s) for s in sorted(ss))
print(f"{str(dataset.name):<25s} {ss}")gives the following output
dataset n_cells
----------------------------------------------
in-L1 105
in-L23 156
in-L4 113
in-L5 111
in-L6 62
lida-in-merged 423
lida-in-merged-bc-merged 423
lida-janelia-L5 58
pc-L2 40, 41, 43
pc-L3 44
pc-L4 89
pc-L5 160
pc-L6 125, 128, 129
Context
When looking at the results of the
performance tableit looks like, for some datasets, thechance accuracydepends on the choice of the neurite!So either there is a bug in how we compute
change accuracyor theground truthsused for each experiment are different. After some inspection, it looks like this is indeed due to differentground truths, because runninggives
Actions