[
https://issues.apache.org/jira/browse/MADLIB-1446?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Frank McQuillan reassigned MADLIB-1446:
---------------------------------------
Assignee: Advitya Gemawat
> DL: Hyperband phase 2 - generate MST table
> -------------------------------------------
>
> Key: MADLIB-1446
> URL: https://issues.apache.org/jira/browse/MADLIB-1446
> Project: Apache MADlib
> Issue Type: New Feature
> Components: Deep Learning
> Reporter: Frank McQuillan
> Assignee: Advitya Gemawat
> Priority: Major
> Fix For: v1.18.0
>
>
> Python code to do some version of this is in
> https://github.com/apache/madlib-site/blob/asf-site/community-artifacts/Deep-learning/automl/hyperband-diag-cifar10-v1.ipynb
> in methods called `setup_full_schedule()` and `create_mst_superset()` +
> combine with the random search function from
> https://www.pivotaltracker.com/story/show/173692930
> **Story***
> Generate the MST table and do input validation on input params (to the extent
> possible without implementing the whole method). It does not do the whole
> hyperband method. The proposed interface:
> {code}
> madlib_keras_automl(
> source_table, -- input
> model_output_table, -- output
> model_selection_table, -- output
> model_arch_table, -- input
> model_id_list,
> compile_params_grid,
> fit_params_grid,
> automl_method, -- new params vvv
> automl_params
> random_state, -- optional -- from
> generate model configs vvv
> object_table -- optional
> use_gpus, -- optional -- from fit
> multiple vvv
> validation_table, -- optional
> metrics_compute_frequency, -- optional
> name, -- optional
> description -- optional
> )
> {code}
> Here are the output tables:
> (1)
> <model_output_table>
> Same as model output table in
> https://madlib.apache.org/docs/latest/group__grp__keras__run__model__selection.html
> e.g., for R=81 and n=3 will have 81+27+9+6+5 rows
> (2)
> <model_output_table>_summary
> Same as model output table summary in
> https://madlib.apache.org/docs/latest/group__grp__keras__run__model__selection.html
> will have 1 row + add the following columns at the bottom, i.e., right side
> of the table:
> {code}
> use_gpus BOOLEAN
> e.g., TRUE -- this is missing from summary table from
> before
> automl_method TEXT e.g.,
> 'hyperband'
> automl_params_names TEXT[] e.g.,
> {'R', 'eta', 'skip_last' }
> automl_params_vals TEXT[] e.g.,
> {'81', '3', 'TRUE'} -- note this needs to be text array since mixed
> types of autoML params
> {code}
> (3)
> <model_output_table>_info
> Same as model output table info in
> https://madlib.apache.org/docs/latest/group__grp__keras__run__model__selection.html
> e.g., for R=81 and n=3 will have 81+27+9+6+5 rows + add the following columns
> at the bottom, i.e., right side of the table:
> {code}
> s INTEGER
> "Bracket number" e.g.,
> 4
> i INTEGER
> "Depth in bracket model trained to" e.g., 3
> {code}
> (4)
> <model_selection_table>
> Same as model selection table in
> https://madlib.apache.org/docs/latest/group__grp__keras__setup__model__selection.html
> e.g., for R=81 and n=3 will have 81+27+9+6+5 rows
> (5)
> <model_selection_table>_summary
> Same as model selection table in
> https://madlib.apache.org/docs/latest/group__grp__keras__setup__model__selection.html
> **Acceptance**
> 1) For `R=81, eta=3` check that it creates the correct MST tables
> <model_selection_table> and <model_selection_table>_summary
> 2) Set `skip_last =1` and check that it creates the correct MST tables
> 3) Try multiple other values to see if produces the correct schedule
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
