[MADLIB-1304] Figure out image input shapes allowed for keras_fit()

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: v1.16
• Component/s: Module: Neural Networks
• Assignee: Unassigned
• Reporter: Frank McQuillan
• Created: 07/Mar/19 19:24
• Updated: 23/Apr/19 17:20
• Resolved: 23/Apr/19 17:20

Description

Context/assumption

Assume the user has the input shape correct to match the model architecture. It's up to them.

Story

Do not re-shape image inputs.

Add a validation check to fit() that input shape = model architecture expected shape (not 100% guarantee) but will help to catch many errors.

Options

channel first, e.g., (3, 224, 224)
channel last, (224, 224, 3)
flattened
other? YES - all shapes supported, we will just mini-batch and pass it along

Questions

1) Q: which input formats to support?
A: all

2) Q: What input validation to do?
A: Add a validation check in fit() to ensure input shape = model architecture expected shape (not 100% guarantee) but will help to catch many errors.

3) Q: Does current code in DL mini-batch pre-processor pack correctly as is and pass it thru?
A: Yes but need to do some testing