Model_T and Model_Y in DML/ORF

[ziweicong]

Hi! I have a question on model_T and model_Y in DML/ORF. I notice that, by default, model_T and model_Y use Lasso in scikit-learn, while I'm thinking to handle the high-dimensional confounding factors, W, in a nonlinear way, such as using a neural network to fit model_T and model_Y. Theoretically, it's workable (just as mentioned in papers). But in practice, is that possible to train a neural network in Keras and then use it to fit model_T and model_Y?

I try to implement this idea, but get some problems. After trained my neural network in Keras, I saved it using joblib.dump, and then use it to fit model_T and model_Y while calling ORF. However, Python returns some errors:
AttributeError: 'History' object has no attribute 'predict'
and
joblib.externals.loky.process_executor.BrokenProcessPool: A result has failed to un-serialize. Please ensure that the objects returned by the function are always picklable.

I wonder whether Keras models can be used to fit model_T and model_Y? It looks like simply dumping the model from Keras doesn't work. Maybe I need to write a class, like those in scikit-learn, to make Keras models compatible with ORF/DML? Any ideas? Thanks!

[vasilismsr]

You would need to wrap a KerasModel with a class that looks like a scikit-learn estimator, i.e. that has a fit and predict method.

You can see an example of such a wrapper that we also used in the dmliv prototype here:
https://github.com/microsoft/EconML/blob/72a7022d99b92149dca8a09735923d1952a102c2/prototypes/dml_iv/deep_dr_iv.py#L9

where the input h to the model should be a KerasModel, e.g. it could be sth like:

h=keras.Sequential([keras.layers.Dense(1000, activation='relu'),
                                            keras.layers.Dropout(0.17),
                                            keras.layers.Dense(1)])

Then you can simply set model_y=_KerasModel(h), model_t=_KerasModel(h), where _KerasModel is the wrapper class.

[congcongruc]

@vasilismsr Thanks! I'm working on the problem now, using a similar approach. Hope to solve it soon. Thank you!

[vasilismsr]

Consider also using skorch:
https://skorch.readthedocs.io/en/stable/
It offers a scikit-learn interface for any pytorch neural net model. So you can directly use the skorck models as model_y and model_t in DML.

[congcongruc]

@vasilismsr Great! That's more convenient than wrapping a Keras model! I definitely would like to try.

Previously I also consider Scikit-learn MLPRegressor, a simple Multi-layer Perceptron model. But this implementation is not intended for large-scale data, as Scikit-learn doesn't offer GPU support. Its framework also doesn't allow much flexibility to build deep learning architectures.
https://scikit-learn.org/stable/modules/generated/sklearn.neural_network.MLPRegressor.html

I'm now trying to use ORF, with Model_Y and Model_T being neural network models. The implementation would probably be similar to what we do in DML. I'm also interested in building a multi-input neural networks to fit Y and T (e.g., as seen in the attached graph). For example, you may want to train the main input (e.g., text data) and the auxiliary input (e.g., time trend) with different architecture.

Therefore, the neural network will require multiple inputs, with each goes through a branch and then concatenated into one output. I'm trying to implement this idea on ORF's Model_T and Model_Y. I guess this can be done by splitting W into multiple parts in the code?

I hope it's clear. Thanks!

[congcongruc]

Hi!

My problem has been solved using Pytorch. Pytorch allows much flexibility in customizing neural networks (even with multiple inputs). But to make the PyTorch models compatible with ORF's main code, there is a need to reshape the inputs, X, T, and Y. Put it simply, Pytorch requires (and also produces) data formats that slightly different from sklearn models.

[vasilismsr]

If you have a working template code, would be great if you could paste it in this discussion for future users that want to use the same. Also if you want to contribute to the main library a pytorch extension let us know.

[congcongruc]

Sure, I would like to share when the code is ready! Let's keep in touch!

[lauracosgrove]

Hi @congcongruc, thanks for posting this issue. I'm interested in doing a similar thing and would really benefit from reading your template code, if it's available!