Using an embedding as a feature in a seperate model

[Marla Willemse]

I am attempting to reduce the dimensionality of a categorical feature by extracting an embedding layer from a neural net and using it as an input feature in a seperate XGBoost model. 

An embedding layer has the dimensions (nr. unique categories, chosen output size). How can it be concatenated to the continuous variables in the original training data with the dimensions (nr. observations, nr. features)? Since each row in the embedding array represents a category, I assume that we can inner join the embedding array to our training observations on category, but how can we be certain which category is represented by a given row in the embedding array?

If the embedding output size is 10, will this constitute 10 different features when used as input data in a separate model? 

Does anyone perhaps know of a code example where embeddings are used as inputs to a seperate model of a different type?

Thanks in advance!

[Yaoshiang Ho]

One of the winners of this kaggle competition took the output of a convnet and put it into xgboost, which may be helpful

https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection

The basic solution to what you want would be to take an intermediate featuremap of shape NHWC (maybe N 64 64 128) and flatten it to NF where F = HxWxC. Now you have F features per image. 

(If it's the final featuremap before passing into the final dense classifier, a more modern solution would be to do a globalaveragepool along the H and W dims, resulting in a featuremap of NF where F=C).

Forward prop on that, unbatch, and now you have F features per image to pass to your GBDT. 

Y

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/24179c86-3e03-4973-aaaa-f450a5bf1cdcn%40googlegroups.com.

[Marla Willemse]

Much thanks for your answer, Yaoshiang!

I am still not sure how to map rows in the embedding layer back to categories/ labels in the categorical column from which the embedding layer was derived. I am attempting to train a simple regression model. 

I have tried to ask my question more clearly on Stack Overflow: https://stackoverflow.com/questions/67610396/how-to-use-a-learned-embedding-layer-from-a-keras-ann-as-an-input-feature-in-an 

[Yaoshiang Ho]

Ah, sorry, I made the assumption you were training an XGBoost layer on top of a convnet. What I suggested above doesn't apply to your problem of ... sorry for misleading you here!

Y

[Nguyễn Anh Hải]

Please help I am getting this error when running speaker_identification

Vào Th 5, 20 thg 5, 2021 vào lúc 03:36 Yaoshiang Ho <yaos...@gmail.com> đã viết:

To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/CAK0BDX0PyM7zbYCmoBPrMCbyHMHHUxKiijZ4tJVnrAz%2BmJ3qaQ%40mail.gmail.com.

[Lance Norskog]

Added an answer. 

To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/CAAwGRFF5wc2-22BRxEEOJs2MuO8o-6oZOdchvQJ5PMe1GCfHXA%40mail.gmail.com.

--

Lance Norskog
lance....@gmail.com
Redwood City, CA

[Madiouni Mohsen]

Hi 

i am working with a dataset that have multitarget.

can one single MLP predict all the outputs or should i make for every target a custom  NN ?

[morteza mohammady]

When you have multi target, you need to add multiple output layer for each target.

For example, you want to recognize face skin color and hair color. Assume you have 3 conv layers and two parallel softmax layers; one for skin color and another hair color.  The input of both softmax layers is the output of the last convolution layer

--

You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.

To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/5854b793-24fc-4d4a-8621-eb4e7d805ce4n%40googlegroups.com.