[jira] [Created] (MADLIB-1375) ParamSearch: Add utility function for generating model selection table

["Yuhao Zhang (JIRA)" <ji...@apache.org>]

Yuhao Zhang created MADLIB-1375:
-----------------------------------

             Summary: ParamSearch: Add utility function for generating model selection table
                 Key: MADLIB-1375
                 URL: https://issues.apache.org/jira/browse/MADLIB-1375
             Project: Apache MADlib
          Issue Type: New Feature
          Components: Deep Learning
            Reporter: Yuhao Zhang
             Fix For: v1.17


As part of this JIRA, we create a utility function to generate the model selection table.
The user, inputs the model_arch_table, sets of model_arch_ids, compile_params, and fit_params. The utility function creates a table with a combination of these three inputs, each with a distinct `MST_KEY`.

*API* 
{code}
load_model_selection_table(
model_arch_table, 
model_selection_table, -- output table name
ARRAY[] model_arch_id, -- Assuming there is only 1 model_arch_table and all model_arch_id are from that table
ARRAY[] compile_params,
ARRAY[] fit_params
)
{code}

*Output Model Selection Table
*
|MST_Key (sequence)| model_arch | model_arch_id |compile_params |fit_params| 
|----|----|----|----|----|
|1|model_arch_library|1|C1|F1|
|2|model_arch_library|2|C2|F2|
|3|model_arch_library|3|C3|F3|

*Acceptance*
1. Validate model_arch_table exists and model_selection_table does not exist
2. The utility function should error if the user inputs an invalid model_arch_id
3. The utility function should error if user inputs invalid compile_param/fit_param
4. The utility function should deduplicate model_arch_id, compile_param, fit_param if passed in

*Notes/open questions
1. We may want to offer other options later for automatically generating the `model_selection_table` (e.g., log, random)
2. How to reconcile with Keras and other libraries that have param selection capability? We are working at a higher level here, but can we take advantage nonetheless?



--
This message was sent by Atlassian JIRA
(v7.6.14#76016)