Frank McQuillan created MADLIB-1389:
---------------------------------------
Summary: Transfer learning for multi-model
Key: MADLIB-1389
URL: https://issues.apache.org/jira/browse/MADLIB-1389
Project: Apache MADlib
Issue Type: New Feature
Components: Module: Neural Networks
Reporter: Frank McQuillan
Fix For: v1.17
Context
The transfer learning workflow for 1.17 will be the same as 1.16. It means user
needs to update the model architecture table with the weights to be used for
initialization. If they are NULL, then Keras default initialization will be
used (random, or perhaps what is specified in the model architecture which I
think there might be options for initialization in model architecture).
Story
I think the only bit that is missing currently is to check the model table to
see if there are any weights there, and if there are, to use them for
initialization.
Acceptance
1) Train a model with 4 MSTs and plot the loss/accuracy curves. Use 2 MSTs for
one model architecture and 2 MSTs for a second model architecture. Perhaps use
CIFAR-10 dataset.
2) Copy model weights over to the model architecture table for 2 of the models
from step #1 (not all 4).
3) Train the same 4 models as #1 and plot the loss/accuracy curves. Check that
the 2 transfer learning cases pick up from where they left off, and that the
other 2 start from scratch (due to random initialization).
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
