Annotate intertrack edges in the ctc get_edge_errors function
#54
Conversation
| induced_graph = comp_graph.get_subgraph( | ||
| comp_graph.get_nodes_with_flag(NodeAttr.TRUE_POS) | ||
| ).graph | ||
|
|
||
| comp_graph.set_edge_attribute(list(comp_graph.edges()), EdgeAttr.FALSE_POS, False) | ||
| comp_graph.set_edge_attribute(list(comp_graph.edges()), EdgeAttr.TRUE_POS, False) |
There was a problem hiding this comment.
@msschwartz21 I don't think I fully understand why we are not annotating true positive edges? You mention in the description that we can't rely on the assumption that TP and WS edges are mutually exclusive, but I think we can? I don't think a TP edge should ever be annotated with WS.
There was a problem hiding this comment.
My stab at explaining this, that Morgan can elaborate on at her leisure: I admit that I am always confused byWS edges; however, it is my understanding that if there is a division and we correctly detect one of the edges to a daughter but not the other, the one we correctly detected would be labeled WS - correct? I would consider that a TP - we correctly identified the edge! Just because we missed the other daughter edge (a FN edge), that doesn't make the correctly recovered edge incorrect. In that way, something that is a WS edge in the CTC metric could be considered a TP elsewhere, making them not mutually exclusive.
There was a problem hiding this comment.
^ Caroline's description captures my thoughts perfectly
I think it was a bug on my end, I don't see anything wrong with what we're doing here and the additional edges (in main, not here) shouldn't have been intertrack edges anyway. So just ignore this hehe |
|
@msschwartz21 Are we good to merge now? |
|
I want to respond to Draga's question about true positive edges at some point, but it will take some additional thought on my part and shouldn't be a blocker to merging. |
I took a stab at it, and you can follow up whenever you have a chance. I'll go ahead and merge then! Thanks :) |
This PR moves the annotation of intertrack edges into
get_edge_errorsbased on the assumption that any edge connecting a parent to a daughter cells is an intertrack edge. This closes #45.A couple additional changes that emerged:
get_edge_errorsexplicitly checks that node errors have been annotated and computes them if they are missing (sinceget_edge_errorsrequires access to true positive nodes)get_edge_errors. The CTC metrics do not use the true positive annotation. Currently true positive edge and wrong semantic edge were considered mutually exclusive which Caroline and I both thought was a potentially dangerous assumption for someone who wasn't operating with CTC metrics. Without the true positive edges explicitly annotated, anyone who wants to count them can impose their own assumptions about what merits a true positive edge.