[
https://issues.apache.org/jira/browse/MADLIB-1336?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16854969#comment-16854969
]
Ekta Khanna commented on MADLIB-1336:
-------------------------------------
Merged as part of PR: [https://github.com/apache/madlib/pull/385]
> Some compile and fit parameters not being passed properly to Keras
> -------------------------------------------------------------------
>
> Key: MADLIB-1336
> URL: https://issues.apache.org/jira/browse/MADLIB-1336
> Project: Apache MADlib
> Issue Type: Bug
> Components: Module: Neural Networks
> Reporter: Frank McQuillan
> Priority: Major
> Fix For: v1.16
>
>
> I suggest that we support the optional / mandatory rules of the underlying
> Keras compile and fit parameters.
> (1) I did not use a metrics parameter and it failed:
> {code}
> SELECT madlib.madlib_keras_fit('mnist_train_packed', -- source_table
> 'model1', -- model
> 'dependent_var', -- dependent_varname
> 'independent_var', -- independent_varname
> 'model_arch_library', -- model_arch_table
> 2, -- model_arch_id
> $$ loss='categorical_crossentropy',
> optimizer='adam' $$, -- compile_params
> $$ batch_size=128, epochs=4 $$, -- fit_params
> 1, -- num_iterations
> FALSE, -- use_gpu
> NULL, -- validation_table
> 'Frank', -- name
> 'A test model' -- description
> );
> {code}
> produces
> {code}
> InternalError: (psycopg2.InternalError) plpy.SPIError: KeyError: 'metrics'
> (plpython.c:5038) (seg0 slice1 10.128.0.41:40000 pid=31836) (plpython.c:5038)
> DETAIL:
> Traceback (most recent call last):
> PL/Python function "fit_transition", line 6, in <module>
> return madlib_keras.fit_transition(**globals())
> PL/Python function "fit_transition", line 388, in fit_transition
> PL/Python function "fit_transition", line 71, in compile_and_set_weights
> PL/Python function "fit_transition", line 241, in compile_model
> PL/Python function "fit_transition", line 144, in
> parse_and_validate_compile_params
> PL/Python function "fit_transition", line 247, in
> validate_compile_param_types
> PL/Python function "fit_transition"
> CONTEXT: Traceback (most recent call last):
> PL/Python function "madlib_keras_fit", line 21, in <module>
> madlib_keras.fit(**globals())
> PL/Python function "madlib_keras_fit", line 157, in fit
> PL/Python function "madlib_keras_fit"
> [SQL: "SELECT madlib.madlib_keras_fit('mnist_train_packed', --
> source_table\n 'model1', -- model\n
> 'dependent_var', -- dependent_varname\n
> 'independent_var', -- independent_varname\n
> 'model_arch_library', -- model_arch_table\n
> 2, -- model_arch_id\n $$
> loss='categorical_crossentropy', optimizer='adam' $$, -- compile_params\n
> $$ batch_size=128, epochs=4 $$, -- fit_params\n
> 1, -- num_iterations\n
> FALSE, -- use_gpu\n NULL, --
> validation_table\n 'Frank', -- name\n
> 'A test model' -- description\n
> );"]
> {code}
> (2) Errors on loss parameter which is optional:
> {code}
> madlib=# SELECT madlib.madlib_keras_fit('mnist_train_packed', -- source_table
> madlib(# 'model1', -- model
> madlib(# 'dependent_var', -- dependent_varname
> madlib(# 'independent_var', --
> independent_varname
> madlib(# 'model_arch_library', --
> model_arch_table
> madlib(# 2, -- model_arch_id
> madlib(# $$ optimizer='adam' $$, --
> compile_params
> madlib(# $$ $$, -- fit_params
> madlib(# 1, -- num_iterations
> madlib(# FALSE, -- use_gpu
> madlib(# NULL, -- validation_table
> madlib(# 'Frank', -- name
> madlib(# 'A test model' -- description
> madlib(# );
> INFO: Model architecture size: 1KB
> CONTEXT: PL/Python function "madlib_keras_fit"
> INFO: Model state (serialized) size: 2MB
> CONTEXT: PL/Python function "madlib_keras_fit"
> ERROR: plpy.SPIError: KeyError: 'loss' (plpython.c:5038) (seg0 slice1
> 10.128.0.41:40000 pid=1039) (plpython.c:5038)
> DETAIL:
> Traceback (most recent call last):
> PL/Python function "fit_transition", line 6, in <module>
> return madlib_keras.fit_transition(**globals())
> PL/Python function "fit_transition", line 388, in fit_transition
> PL/Python function "fit_transition", line 71, in compile_and_set_weights
> PL/Python function "fit_transition", line 241, in compile_model
> PL/Python function "fit_transition", line 136, in
> parse_and_validate_compile_params
> PL/Python function "fit_transition", line 187, in parse_loss
> PL/Python function "fit_transition"
> CONTEXT: Traceback (most recent call last):
> PL/Python function "madlib_keras_fit", line 21, in <module>
> {code}
> (3) All fit params are optional so should be able to pass NULL
> {code}
> madlib=# SELECT madlib.madlib_keras_fit('mnist_train_packed', -- source_table
> madlib(# 'model1', -- model
> madlib(# 'dependent_var', -- dependent_varname
> madlib(# 'independent_var', --
> independent_varname
> madlib(# 'model_arch_library', --
> model_arch_table
> madlib(# 2, -- model_arch_id
> madlib(# $$ optimizer='adam' $$, --
> compile_params
> madlib(# NULL, -- fit_params
> madlib(# 1, -- num_iterations
> madlib(# FALSE, -- use_gpu
> madlib(# NULL, -- validation_table
> madlib(# 'Frank', -- name
> madlib(# 'A test model' -- description
> madlib(# );
> ERROR: TypeError: cannot concatenate 'str' and 'NoneType' objects
> (plpython.c:5038)
> CONTEXT: Traceback (most recent call last):
> PL/Python function "madlib_keras_fit", line 21, in <module>
> madlib_keras.fit(**globals())
> PL/Python function "madlib_keras_fit", line 127, in fit
> PL/Python function "madlib_keras_fit"
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
