[MADLIB-1332] DL: Support mini-batched validation data for fit/evaluate

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: v1.16
• Component/s: Deep Learning
• Assignee: Jingyi Mei
• Reporter: Domino Valdano
• Created: 26/Apr/19 19:15
• Updated: 06/Jun/19 00:40
• Resolved: 06/Jun/19 00:40

Description

Currently, keras_evaluate() is implemented by calling internal_keras_evaluate() as a UDF.  This requires the validation table passed to keras_fit() to be in a format with only 1 image per row, even though the training table is in a different format, with a batch of images in every row.  This is potentially confusing and cumbersome for users to deal with, and based on some preliminary testing it seems that passing only 1 image at a time to keras_evaluate() is also slowing down performance.

We can solve this by converting internal_keras_evaluate() into a UDA, so that it runs on a minibatched validation table in the same form as the training table.

 

Tasks:

• Convert the internal_keras_evaluate UDF to a UDA and perform weighted averaging of loss and accuracy.
• Since x and y will now be minibatched, we don't need to add another dimension to x and y np arrays in internal_keras_evaluate.
• Compare UDF to UDA and verify that the UDA results in a speed improvement