Add PyTorch backend for MF

[hieuddo]

Description

• add PyTorch backend for MF
• add TorchMF to examples/biased_mf.py
• MF minor fix: it should be either user_idx or item_idx is unknown?
  

  cornac/cornac/models/mf/recom_mf.pyx

  Line 269 in b44e19c

  if not self.knows_user(user_idx) or self.knows_item(item_idx):

Checklist:

• I have added tests.

[hieuddo]

Results obtained from examples/biased_mf.py:

	MAE	RMSE	NDCG@10	Recall@10	Train (s)	Test (s)
GlobalAvg	0.9408	1.0835	0.0085	0.0023	0.0000	8.1069
MF	0.6910	0.8470	0.1058	0.0335	0.4068	36.5108
TorchMF	0.6947	0.8590	0.1014	0.0308	119.5144	10.5325

	MAE	RMSE	NDCG@10	Recall@10	Train (s)	Test (s)
GlobalAvg	0.9373	1.0797	0.0088	0.0026	0.0000	9.0318
MF	0.6911	0.8473	0.0976	0.0318	0.4288	32.9461
TorchMF	0.6962	0.8608	0.0977	0.0313	120.6941	11.7570

[lthoang]

do you mean ies (item embeddings) here?

[tqtg]

@hieuddo I did some code refactoring to merge them into one MF model with different options of backends. The results appear to be similar to yours.

TEST:
...
           |    MAE |   RMSE | NDCG@10 | Recall@10 | Train (s) | Test (s)
---------- + ------ + ------ + ------- + --------- + --------- + --------
GlobalAvg  | 0.9426 | 1.0868 |  0.0093 |    0.0036 |    0.0000 |   6.3037
MF-cpu     | 0.6914 | 0.8494 |  0.0899 |    0.0293 |    0.1364 |   9.3481
MF-pytorch | 0.7023 | 0.8579 |  0.1422 |    0.0440 |   68.6235 |   9.4482

[hieuddo]

Looks great.

One minor, though: should we unify formatter settings to reduce changes due to different formatters?

I'm currently using black, autopep8, and isort altogether, with black as the default formatter. I think black is the priority because very few configurations are allowed, which makes code look the same every time; while autopep8 ignores some parts and only formats where the code doesn't follow pep8 style.

[tqtg]

@hieuddo I'm also using black, and we do check PEP8 style(see Link with flake8 step). While trying to enforce convention for our core codebase, we're a bit more flexible with models to ease the contribution.