Extract embedding from my model for other use.

[jjkai]

Hi I have question on how can I extract the embedding from my model for my other downstream tasks. Say for example, running k-means clustering.

Context: I am training a text classification model following the example here: https://ludwig.ai/latest/examples/text_classification/
However, instead of getting the prediction result for each of the records, I am also interested in extracting the embedding for each record: the output from the last fully connected layer before making the prediction. That way I can try to run some other analysis to see what information the model is capturing. As simple as running a k-means clustering to see if each cluster contains similar data records.

Any advise and feedback are appreciated. Thank you.

[tgaddair]

Hey @jjkai, good question. This is possible in Ludwig using the CLI command ludwig collect_activations as described here. Unfortunately we don't have a lot of great examples of this (we should add some!) but we do have an example test here that demonstrates usage:

https://github.com/ludwig-ai/ludwig/blob/master/tests/integration_tests/test_collect.py#L91

Let me know if you have any follow-up questions or suggestions for how we can improve the docs / features around this!

[jjkai]

Thank you! I will give it a try and let you know if I have further questions/suggestions.

[jjkai]

Hi @tgaddair I tried to run the collect_activations function and it generates three files: 'tag__last_hidden.npy', 'tag__logits.npy', and 'tag__projection_input.npy', which I believe is the expected behavior. So the 'tag__logits.npy' looks pretty straightforward to me should be the final predicted logits of an input. But I am not super certain about 'tag__last_hidden.npy' and 'tag__projection_input.npy'. I am guessing 'tag__last_hidden.npy' is probably the node weights for the last layer? And 'tag__projection_input.npy' probably is the embedding I am looking for? But this is all my speculation, can you please verify or provide more detail?

One other issue I am encountering is that when I run the collect_activations function with a dataset say 200 records, sometimes the returned files 'tag__last_hidden.npy' and 'tag__projection_input.npy' would contain less than 200 records. This is likely the sign that I am done something incorrectly. In addition, when looking into the values of the returned records (each is a vector with 256 values), I found that the majority of them are 0. I am attaching an image below.

Not exactly sure what went wrong in the process. Maybe as you suggested, a more concrete example would be great and super helpful. But in the meantime, I appreciate if you can share some insights or maybe know what the issue is already.

Thanks!