Classifier app does not know about post binning

[marcverhagen]

Because

The classifier only works well if there is no post binning. It gets its labels from

app-swt-detection/modeling/train.py

Lines 268 to 274 in 468aa2b

	def get_final_label_names(config):
	base = FRAME_TYPES
	if config and "post" in config["bins"]:
	base = list(config["bins"]["post"].keys())
	elif config and "pre" in config["bins"]:
	base = list(config["bins"]["pre"].keys())
	return base + [modeling.negative_label]

And those labels are added to the Prediction instances. If there is post binning than the labels will be out of sink with the Prediction because the scores in the Prediction are for the labels after pre binning, not post binning.

Changes needed are to Classifier.process_video() which hands over the wrong labels to the prediction, but mostly to the Stitcher, which now does not know what scores to pick.

Note to self: post binning as implemented in train.py does not involve adding probabilities on post-binned labels, instead it picks the highest score. It should not be hard to experiment with both in the stitcher.

Done when

No response

Additional context

No response

[marcverhagen]

Fixed in #57

[marcverhagen]

@keighrim do you know why this was reopened? CLAMS Bot gone rogue?

[keighrim]

It look like the bot assigned this issue to you when you pushed 14-clamsapp-xxx branch, and as part of assigning action it reopened it. (https://github.com/clamsproject/app-swt-detection/actions/runs/7821239106/job/21337739873)

[keighrim]

No that was issue 14, not this one.

[keighrim]

Aha, you also pushed 46-postbinning branch around the same time (and deleted immediately), and bot caught that.

[marcverhagen]

Ah, I did accidentally do a push --all in stead of a push --tags and that included many branches so that could be it. I was the one who went rogue.