Does MxNet have anything like TensorFlow Feature Columns?

Discussion
1

Hello all,

I am quite new to the machine learning ecosystem, and have started to learn this is as a hobby and maybe a career. I thing I am not sure what the best practice is for keeping track of feature names to ndarray indices. For example, I have a 1D feature vector X of say height and age. How do I best keep track of the fact that X[0] is height and X[1] is age?

In TensorFlow there seems to be a something called “Feature Columns”. https://www.tensorflow.org/guide/feature_columns that seems to handle the mapping from feature name to index for you. Like you can feed it a dict/map with your data.

In pytorch, there does not seem to be such a thing. https://github.com/pytorch/pytorch/issues/22610

How about in mxnet? I cannot seem to find something like this.

Is is the end user’s responsibility of maintaining a string int bidirectional map of feature name to and from ndarray index for your Module?

Thank you,

2

Hi @rko,

I don’t think MXNet has such a feature. As you said, the user should be responsible of maintaining the maps, as there wouldn’t be any benefit of doing it in MXNet since it is easily to implement in vanilla Python.