Transfer Learning with Structured Data Question

Part 2 & Alumni (2018)
1

Does anyone have experience with this?

I have one dataset with about 1 million rows and another with 210,000.

I have successfully built a 3 dense layer model (excluding embeddings of which i have many) that does a really good job of predicting my 16 continuous outputs.

I’m considering three approaches.

  1. Add 2-3 dense layers to the end of the existing model
  2. retrain the existing model with very low learning rates
  3. don’t retrain at all and push the 210,000 row dataset through the existing model with no changes.

I’m open to suggestions or ideas, thanks guys.

2

quick update:

tried method 3 and unsurprisingly is doesn’t work very well, although not terrible.

3

Hi Whamp, did you ever figure this out ?

4

Unfortunately I got pulled onto something else and never got a chance to experiment. However I’m actually now getting a chance to revisit this and will update here if I make any progress.