[jira] [Closed] (MADLIB-1441) DL - Add support for custom loss functions

["Frank McQuillan (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/MADLIB-1441?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Frank McQuillan closed MADLIB-1441.
-----------------------------------
    Resolution: Fixed

https://github.com/apache/madlib/pull/503

> DL - Add support for custom loss functions 
> -------------------------------------------
>
>                 Key: MADLIB-1441
>                 URL: https://issues.apache.org/jira/browse/MADLIB-1441
>             Project: Apache MADlib
>          Issue Type: New Feature
>          Components: Deep Learning
>            Reporter: Ekta Khanna
>            Priority: Major
>             Fix For: v1.18.0
>
>
> Keras supports custom loss functions to be passed in as compile_params to the fit function. For MADlib deep_learning module to support custom functions, we have the following functions are affected:
> 1. madlib_keras_fit()
> 2. madlib_keras_fit_multiple_model()
> 3. madlib_keras_evaluate()
> Currently the compile_param passed in are like
>  {{optimizer=SGD(lr=0.01, decay=1e-6, nesterov=True), loss='categorical_crossentropy', metrics=['mae']}}
>  For a custom function we would just pass in as follows:
>  {{optimizer=SGD(lr=0.01, decay=1e-6, nesterov=True), loss=custom_fn1, metrics=['mae']}} 
>  
> object table definition with a dill object of custom functions is as follows:
> ||id||name||description||function||
> |1|custom_fn1|test function|\x80026...|
> |1|custom_fn2|test function2|\x8ab26...|
>  
> For fit_multiple, we want to support multiple configs in mst_table with a mix of custom and keras loss functions:
> ||mst_key||model_id||compile_params||fit_params||
> |2|1|loss='custom_fn1', optimizer=....|batch_size=16, epochs=1|
> |3|1|loss='custom_fn2', optimizer=....|batch_size=16, epochs=1|
> |1|1|loss='categorical_crossentropy', optimizer=....|batch_size=16, epochs=1|
>  
> h4. API Changes
>  # *Fit*: add a new optional param for passing in the object_table name
>  object_table (optional) VARCHAR: Name of the table containing Python objects in the case that custom loss functions or custom metrics are specified in the parameter `compile_params
> {code:java}
> madlib_keras_fit(
>     source_table,
>     model,
>     model_arch_table,
>     model_id,
>     compile_params,
>     fit_params,
>     num_iterations,
>     use_gpus,
>     validation_table,
>     metrics_compute_frequency,
>     warm_start,
>     name,
>     description,
>     object_table  -- new parameter
>     ){code}
>  # *Fit_multiple*: No change to the {{madlib_keras_fit_multiple_model()}} function. Reads object_table information from the model_selection table.
>  A summary table named <model>_summary is also created, which has the following new columns:
> |model_selection_table |Name of the table containing model selection parameters to be tried.|
> |object_table| Name of the object table containing the serialized Python objects for custom loss functions and custom metrics (read from the mst_summary table).|
>  # *Evaluate*: No change to the {{madlib_keras_evaluate()}} function. Reads object_table information from the model table. Output table adds a new column
> |loss_type|Type of loss used that was used in the training step
>  If a custom loss or metric is used, we should give the name of it. Otherwise list the built-in one used|



--
This message was sent by Atlassian Jira
(v8.3.4#803005)