[Enhacemnet] api train support cpu training for mmcv<1.4.4

[EasonQYS]

[Enhance] support training api on cpu

Motivation

support training api on cpu

Modification

add a parameter 'device' and add a if-else in train_module(...)

Checklist

Before PR:

• I have read and followed the workflow indicated in the CONTRIBUTING.md to create this PR.
• Pre-commit or linting tools indicated in CONTRIBUTING.md are used to fix the potential lint issues.
• Bug fixes are covered by unit tests, the case that causes the bug should be added in the unit tests.
• New functionalities are covered by complete unit tests. If not, please add more unit tests to ensure correctness.
• The documentation has been modified accordingly, including docstring or example tutorials.

After PR:

• CLA has been signed and all committers have signed the CLA in this PR.

[CLAassistant]

All committers have signed the CLA.

[jin-s13]

Thanks for your contribution!

Not sure if it is really necessary to support CPU training. It normally takes a lot of time. Haha 😆

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (dev-0.24@18af415). Click here to learn what that means.
The diff coverage is n/a.

❗ Current head cd89d70 differs from pull request most recent head 5435695. Consider uploading reports for the commit 5435695 to get more accurate results

@@             Coverage Diff             @@
##             dev-0.24    #1161   +/-   ##
===========================================
  Coverage            ?   82.67%           
===========================================
  Files               ?      204           
  Lines               ?    16365           
  Branches            ?     2943           
===========================================
  Hits                ?    13529           
  Misses              ?     2092           
  Partials            ?      744           

Flag	Coverage Δ
unittests	82.60% <0.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 18af415...5435695. Read the comment docs.

[EasonQYS]

Thanks for your contribution!

Not sure if it is really necessary to support CPU training. It normally takes a lot of time. Haha 😆

Thanks！
As far as I am concerned, many students (and their family) and developers are just interested at AI and CV, who are not equiped with the PC good enough, but they should not be left out of the consideration. So I recommand to support CPU training.

[ly015]

Thank you very much for your contribution!

In fact, we have just supported CPU training/test in #1157, boosted by new features in mmcv 1.4.4.

Nevertheless, I think it would be nice to also support for earlier mmcv version. So would you mind modifying this PR and adding an mmcv version check to determine whether to use MMDataParallel on CPU? Here is a reference.

[EasonQYS]

Thank you very much for your contribution!

In fact, we have just supported CPU training/test in #1157, boosted by new features in mmcv 1.4.4.

Nevertheless, I think it would be nice to also support for earlier mmcv version. So would you mind modifying this PR and adding an mmcv version check to determine whether to use MMDataParallel on CPU? Here is a reference.

@ly015
Thank you to tell me that, and I am glad to help!

However, there is one thing I am confused, if I use just use a single cpu to train the model, why should I put data to parallel. Is it a standard, or cpu can be more than one pice?

Thank you again!

[ly015]

Thank you very much for your contribution!
In fact, we have just supported CPU training/test in #1157, boosted by new features in mmcv 1.4.4.
Nevertheless, I think it would be nice to also support for earlier mmcv version. So would you mind modifying this PR and adding an mmcv version check to determine whether to use MMDataParallel on CPU? Here is a reference.

@ly015 Thank you to tell me that, and I am glad to help!

However, there is one thing I am confused, if I use just use a single cpu to train the model, why should I put data to parallel. Is it a standard, or cpu can be more than one pice?

Thank you again!

MMDataParallel does not really perform data parallel in CPU mode. It's equivalent to the simple forward of the wrapped PyTorch module. This design is for keeping a unified interface for both CPU and GPU environments.

[EasonQYS]

MMDataParallel does not really perform data parallel in CPU mode. It's equivalent to the simple forward of the wrapped PyTorch module. This design is for keeping a unified interface for both CPU and GPU environments.

Thanks for letting me know!