Multi Task Learning implementation and example

[johann-petrak]

This fixes #408 for me, did some testing with two classification heads so far and most everything works fine: I am getting proper evaluations for each head during training and the inferencer provides a list of two predictions for each instance.

Still to check:

• why do we not get the proper task name for the predictions from the inferencer?
• what if we use one or more per-token heads?

Help with these things to check and with how to best test this in general would be appreciated.

Timoeller edit: fixes #742

[Timoeller]

Incredible. You are doing a little sweep through all of FARM and it is much cleaner than before : ) Thanks a lot @johann-petrak

[tholor]

I had a quick look at your code and I think your changes make sense. Haven't tried if they are breaking anything around NaturalQuestions, but the NQ inference part was broken already before and we will need to refactor this anyways.

why do we not get the proper task name for the predictions from the inference?

I guess you mean "task":"text_classification" in the dict returned by the inferencer? This is set in the head itself here and is rather reflecting the task type: https://github.com/deepset-ai/FARM/blob/master/farm/modeling/prediction_head.py#L413

I believe the order of the predictions should reflect the order of your prediction heads. However, I can totally see that for multiple classification heads (especially with very similar predictions) you want a clear identification in the results. We could probably add a field "task_name" to the dict at the above location in the code by reading from the prediction head's self.task_name.

[johann-petrak]

I believe the order of the predictions should reflect the order of your prediction heads. However, I can totally see that for multiple classification heads (especially with very similar predictions) you want a clear identification in the results. We could probably add a field "task_name" to the dict at the above location in the code by reading from the prediction head's self.task_name.

My main concern is that it is a bit confusing how the keys/parameter names "task", "task_type" and "task_name" are used, ideally it would always be according to the convention that "task" is a task object, "task_name" is the name of a task, and "task_type" is the type, but this is not always the case. (I know I can be a bit too OCD about such things)

Of course hard to change now as it would break backwards compatibility, so I have no opinion :)

Adding a field "task_name" in the prediction would definitely help to immediately realize that "task" is not meant to be the task name. Should we make this part of this PR or have a different issue for it?

[Timoeller]

Hey @johann-petrak I agree that the task setup could be improved - inconsistent use of naming is very confusing to users (and also us, looking at code we wrote a year ago and just shaking heads), so we also appreciate rigorous improvement on this end.

In general tasks is a nested dict looking like:

self.tasks[name] = {
            "label_list": label_list,
            "metric": metric,
            "label_tensor_name": label_tensor_name,
            "label_name": label_name,
            "label_column_name": label_column_name,
            "text_column_name": text_column_name,
            "task_type": task_type
        }

and I like the suggestion to add the tasks name as "task_name" to the prediction heads output. People might want to have n classification heads that would be only distinguishable through this task_name.

So lets add this "task_name" in this PR and improve upon the general naming of tasks in a separate PR?

[Timoeller]

Hey @johann-petrak any progress on the task_name field?

I think the PR is already in good shape and could be merged soon.

[johann-petrak]

Hey @johann-petrak any progress on the task_name field?

I think the PR is already in good shape and could be merged soon.

Sorry, did not forget about this, I have been crazy busy with other stuff and planning to get back this this in the next days.

[johann-petrak]

OK, this is now now ready from my side, the prediction heads for text classifcation, regression and a few others now include the key "task_name" in the formatted prediction result. The key "task" contains the type as before. There is also an example script to illustrate using two text classification heads.

This did not address or test using multiple heads where the heads are of different type, e.g. classification and NER. Probably better to open or re-use a separate issue for that, as it will need some good dataset for testing as well.

[johann-petrak]

Sorry, that was my insanity, that file should not have been added, thanks for spotting! Removed now.

[julian-risch]

Had some problems with executing the benchmark but now the results are here: https://github.com/deepset-ai/FARM/pull/797#commitcomment-51938273 -> Ready to merge

Benchmarks: QA Accuracy

:heavy_check_mark: QA Accuracy Benchmark Passed

	run	f1_change	em_change	tnacc_change	elapsed_change	f1	em	tnacc	elapsed	f1_gold	em_gold	tnacc_gold	elapsed_gold
0	FARM internal evaluation	0.017	0.0001	0.0169	-0.5025	82.6841	78.4722	84.3763	39.4975	82.6671	78.4721	84.3594	40
1	outside eval script	-0.0045	-0.0077	-	0.0436	82.9125	79.8703	-	27.0436	82.917	79.878	-	27
2	train evaluation	0.4809	-0.2106	0.1685	-20.0116	82.6359	78.4469	97.5406	1114.99	82.155	78.6575	97.3721	1135