Explicit class mapping in DetectNet

[gheinrich]

The hard-coded value in the code feels a bit too magical.

[lukeyeager]

I agree. So shouldn't we remove that hardcoded value that you pointed out?

[gheinrich]

I didn't dare to make that change :-)

Those who have been using DetectNet until now might want the default value to remain in place so they can keep using their old model description. Adventurers might want the layer to fail if no explicit mapping was specified. I think it's cleaner to not have the default value at all but I understand that this would break backward compatibility.

[lukeyeager]

But merging this by itself would break things too, right? Since there would no longer be any data labelled as 1?

[gheinrich]

Sorry I don't understand the question. The intention behind this change is to explicitly do the same thing as is done implicitly i.e. assign class "1" to index "0" in the coverage tensor. Why would this break anything?

[lukeyeager]

Oh I see, you're right. This is a noop change that exposes the underlying behavior more explicitly. I like that.

What I don't like is the default behavior when there are no class mappings. Currently, we only pick label 1 if there are no mappings. That seems illogical. We could either (1) only take label 0 or (2) take all labels and do multiclass by default. I'm not sure if (2) would actually work in practice.