[ENH] First transfer of deep learning classifiers and regressors from sktime-dl

[TonyBagnall]

this is my first go at this and open to discussion as to structure. Once we have agreed how to do CNN, the rest of the algorithms will come over quickly. However, there are features here some may dislike, so want to out there early. Points to note

1. Introduces a whole new module called networks. This contains the network structures that are shared by both classifiers and regressors. This is necessary because it was agreed long ago to split the algorithms into regressors/classifiers (rather than have one class with both capabilities, which was my preferred structure)
2. classifiers go in classification/deep_learning, and there is an abstract class there that inherits from BaseClassifier. This intermediate class is there because we can define default DL classifier predict_proba behaviour, and we want an extra abstract method.

[TonyBagnall]

Looks ok, but you need to adhere to the dependency handling via checks, see docs for that.

tried _check_dl_dependencies but it doesn't have a severity argument, so reverted back to _check_soft_dependencies. I'm not sure why there is a separate method for dl dependencies, will surely just be duplication of code, and I know how much you hate that

[fkiraly]

tried _check_dl_dependencies but it doesn't have a severity argument

now it does: #2627

so reverted back to _check_soft_dependencies.

would in-principle be fine by me

I'm not sure why there is a separate method for dl dependencies,

more informative (deep learning specific) error message and developer comfort for an anticipated frequent developer case.

will surely just be duplication of code, and I know how much you hate tha

Agree, could easily be refactored.
Although in this case I would put user and developer comfort above "neat code".

[TonyBagnall]

@fkiraly I believe you have someone working on the porting of sktime-dl, which is great. I would like to get this PR in first in order to fix the structure. I've lost track of what you wanted changing here, are you still blocking it, and if so, what do you want changing?

[fkiraly]

@TonyBagnall, what have you done/decided about the dependency handling?

Also, it's not "me having someone work" on DL. sktime, as a programme & community, is hosting summer projects, one of which is working on this, yes. You are more than welcome to engage with the summer mentees, the meetings are on our calendar, and as CC member you also have a full view of the process (see "updates on GSoC" items).

[TonyBagnall]

@TonyBagnall, what have you done/decided about the dependency handling?

Also, it's not "me having someone work" on DL. sktime, as a programme & community, is hosting summer projects, one of which is working on this, yes. You are more than welcome to engage with the summer mentees, the meetings are on our calendar, and as CC member you also have a full view of the process (see "updates on GSoC" items).

ok, soz my wording offends you, not sure what you mean by "decided about the dependency handling", could you clarify?

[fkiraly]

ok, soz my wording offends you,

I'm not offended, I just wanted to disagree with the (possibly unintentionally) possessive framing.

not sure what you mean by "decided about the dependency handling", could you clarify?

Refers to my change request, in which I said:

Looks ok, but you need to adhere to the dependency handling via checks, see docs for that.

More precisely, I was asking above, for:

• adding informative warning and error messages for developers using the DL functionality
• removing any errors that would be triggered by module import

and with my more recent question, I have asked you for explaining whether and how this has been addressed.

Addressing could mean (a) agreeing with my request and taking some action, or (b) disagreeing and providing a statement why you think you do not need to do it, or (c) thinking something different needs to be done, taking alternative action.

In either case, I would kindly ask for explanation of the action taken and/or rationale.

[TonyBagnall]

I have switched to your _check_dl_dependencies given you went to the trouble of updating it. I think it already complied with the docs on new soft dependencies (warn/error, imports in functions)

[TonyBagnall]

@fkiraly I would like this in before the next release, could you unblock or give specific changes you would like?

[fkiraly]

All good.

[xxl4tomxu98]

is this resolved? e.g. am I able to install the sktime[all_extra] and then all soft dependency are satisfied? including tensorflow?

[xxl4tomxu98]

The problem with sktime-dl is that I am using ARM based Mac
and it won't work when I try to do "conda install sktime-dl"