Summarize model size in MegaBytes [WIP]

[kartik4949]

Fixes # (issue)
#4807
This PR adds a Summarization of the model with estimated total model size.
Estimated Total model size is "input_size + forward/backward pass size + total params size".

e.g
VGG16 with input (1, 1, 28, 28)
Input Size (MBs) --> 0.57
Total Params Size (MBs) --> 528 (e.g downloading weights size)
Forward/Backward Size (MBs) --> 218
Total Estimated Model Size (MBs) --> 747

Testing

1. Test knownet (a simple convolution NN)
   tests knownet model with different input shapes and batch sizes.
   we test the corresponding model sizes with pre calculated model size.
2. Test Different Input Types:
   tests a DummyNetwork with different input types.

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together? Otherwise, we ask you to create a separate PR for every change.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?
• Did you verify new and existing tests pass locally with your changes?
• If you made a notable change (that affects users), did you update the CHANGELOG?

PR review

Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified; Bugfixes should be including in bug-fix release milestones (m.f.X) and features should be included in (m.X.b) releases.

Did you have fun?

Make sure you had fun coding 🙃

[kartik4949]

WIP please suggest how to proceed from here.

[kartik4949]

@awaelchli @rohitgr7 I'm sensing a usage of LRU cache in model summary, as calculating out and input sizes can be redundant,
thoughts?

[codecov]

Codecov Report

Merging #4810 (f477c2f) into release/1.2-dev (24462dc) will increase coverage by 1%.
The diff coverage is 97%.

@@               Coverage Diff                @@
##           release/1.2-dev   #4810    +/-   ##
================================================
+ Coverage               93%     93%    +1%     
================================================
  Files                  153     135    -18     
  Lines                10815   10005   -810     
================================================
- Hits                 10006    9339   -667     
+ Misses                 809     666   -143     

[tchaton]

Would you mind adding some tests ?

[kartik4949]

Would you mind adding some tests ?

@tchaton
sure, wont mind adding tests, you looking for any specific test? :)

[pep8speaks]

Hello @kartik4949! Thanks for updating this PR.

• In the file pytorch_lightning/core/memory.py:

Line 237:101: E203 whitespace before ':'

• In the file tests/core/test_memory.py:

Line 24:5: E731 do not assign a lambda expression, use a def

Comment last updated at 2020-12-07 07:02:33 UTC

[kartik4949]

@justusschock can u suggest tests?

[justusschock]

You could start with simple models like one or two layers, where you know the size / can calculate them by hand and then compare it to the automatically inferred size. Also you could do it once for a more complex model.

[kartik4949]

You could start with simple models like one or two layers, where you know the size / can calculate them by hand and then compare it to the automatically inferred size. Also you could do it once for a more complex model.

Thanks doing the same:)

[kartik4949]

@awaelchli @ananyahjha93 @williamFalcon @Borda I think this PR is ready , need to add some complex models for testing
i.e 1. multiple input model, 2. multi output.
then we are good to go! :)

[Borda]

nice work, just need to finish some minor nits...

[kartik4949]

@Borda dropped the idea of input size, we are good with example_input_array only which will act as custom input_size.

[kartik4949]

@Borda
so made new changes,

1. Dropped input_size, example_array will be used instead
2. Model size shown only in FULL mode.
3. Added new test which checks for nested models.

All Tests are passed :)

[kartik4949]

@awaelchli @justusschock Hey
Currently modelsize = input size + forward/backpass pass size + total params size (MBs)

lets calculate model total trainable params size only
as it will have less complicated and further we can add other sizes too.
what you say ?
if so then I can remove the complex feature.
lets conclude this PR :)
thanks, waiting for some help.

[justusschock]

@kartik4949 Yes, I think that's best. And for the more complicated part, maybe it's better to do this in a separate follow up PR :)

[kartik4949]

@justusschock thanks for responding will remove other sizes and conclude this one :)

[kartik4949]

@justusschock hey, done with changes.
btw i did rebase on 1.15
but pr is for 1.12-dev.
idk how to interpret this :)