feat(skorch): add an inherited class from skorch.NeuralNet that is compatible with PyTorch Frame #375
Conversation
…et that is compatible with PyTorch Frame
for more information, see https://pre-commit.ci
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #375 +/- ##
=======================================
Coverage 93.52% 93.52%
=======================================
Files 124 124
Lines 6456 6456
=======================================
Hits 6038 6038
Misses 418 418 ☔ View full report in Codecov by Sentry. |
|
Sure - on both! |
34j
force-pushed
the
feat/skorch-compatible
branch
from
1706c96
to
0b9426f
Compare
|
@weihua916 Would you mind reviewing if you think this is a good way to implement it? Also, it is strange that mypy in pre-commit does not raise errors, but mypy in CI does. I don't think there is any way to deal with this. |
|
@weihua916 @zechengz @yiweny Would appreciate your review, thank you very much in advance. |
There was a problem hiding this comment.
Thank you!
Do you mind adding unit test similar to https://github.com/pyg-team/pytorch-frame/blob/master/test/gbdt/test_gbdt.py?
|
A kind check-in. Is there any progress here? |
|
No progress, sorry |
for more information, see https://pre-commit.ci
|
@weihua916 Thanks for your patience, I've added tests. |
There was a problem hiding this comment.
Left some comments. Can @weihua916 or @yiweny helps to take a look? Thanks for your contribution!
| def _patch_skorch_support_tenforframe() -> None: | ||
| old_to_tensor = skorch.utils.to_tensor | ||
|
|
||
| def to_tensor(X, device, accept_sparse=False): |
There was a problem hiding this comment.
As well, this is not recommended as skorch does not support typing.
torch_frame/utils/skorch.py
Outdated
| def __init__( | ||
| self, | ||
| # NeuralNet parameters | ||
| module, |
There was a problem hiding this comment.
Also add the typing information here?
There was a problem hiding this comment.
Since skorch does not support typing, doing this is tedious and not recommended. (On the other hand, it would be inconvenient for users if it was **args, **kwargs, so I compromised and made it this way.)
…ough it can be used
|
@weihua916 Removed all tutorials and added |
| # [stype.text_embedded], | ||
| # [stype.numerical, stype.numerical, stype.text_embedded], |
There was a problem hiding this comment.
we don't support these stypes?
There was a problem hiding this comment.
Currently not supported at this time due to lack of time to understand how to use these dtypes.
However, since it probably only require changes in the arguments of the NeuralNet, it should have little trouble extending it in the future.
| if pass_dataset: | ||
| net.fit(dataset) | ||
| _ = net.predict(test_dataset) | ||
| else: | ||
| net.fit(X_train, y_train) | ||
| _ = net.predict(X_test) |
There was a problem hiding this comment.
why don't we take tensor frame? It's also weird to sometimes take dataset and sometimes take data frame.
There was a problem hiding this comment.
- The main purpose of this PR is to allow DataFrame to be DIRECTLY fitted, as shown in examples/sklearn_api.py.
- Since it is unclear how to create a Dataset from a TensorFrame, and if there is a TensorFrame, there should be also a Dataset, which means there is little need to implement this, and even to use skorch as the user might be familiar with deep learning.
- Instead of "sometimes take dataset and sometimes take data frame", both are tested.
There was a problem hiding this comment.
hmm if dataframe is directly fed, it is unclear why we need this feature within pytorch frame.
the whole point of pytorch frame is to materialize data frame into tensor frame, to be processed by pytorch.
There was a problem hiding this comment.
It may not match your purpose but my goal is to use advanced neural networks implemented in pytorch_frame in existing sklearn pipeline.
This PR allows pytorch_frame to be used on top of existing scikit-learn code without having to heavily modify the existing code. Since many people use sklearn Pipeline, especially on Kaggle, it is easy to verify performance changes by changing or assembling the estimator in other people's code to my NeuralNetPytorchFrame. I am convinced that this will be very valuable.
Related: #147
@MacOS Please continue from here if it helps.
Sorry for being so loud, but this took me a whole day, so I would appreciate it very much if you could make me as a co-author if you used this code.