You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@madlib.apache.org by "Frank McQuillan (Jira)" <ji...@apache.org> on 2021/03/24 23:11:00 UTC
[jira] [Created] (MADLIB-1482) DL: metrics compute frequency should
be >=1 only
Frank McQuillan created MADLIB-1482:
---------------------------------------
Summary: DL: metrics compute frequency should be >=1 only
Key: MADLIB-1482
URL: https://issues.apache.org/jira/browse/MADLIB-1482
Project: Apache MADlib
Issue Type: Bug
Components: Deep Learning
Reporter: Frank McQuillan
Fix For: v1.19.0
metrics_compute_frequency should be >= 1 only. Currently it allows neg numbers:
{code}
SELECT madlib.madlib_keras_fit('balanced2_train_packed', -- source table
'model1', -- model output table
'model_arch_library', -- model arch table
1, -- model arch id
$$ loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'] $$, -- compile_params
$$ batch_size=64, epochs=1 $$, -- fit_params
10, -- num_iterations
FALSE, -- use GPUs
'balanced2_test_packed', -- validation dataset
-3, -- metrics compute frequency
FALSE, -- warm start
'Frank', -- name
'Network test run' -- description
);
{code}
produces
{code}
SELECT * FROM model1_summary;
-[ RECORD 1 ]-------------+-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
source_table | balanced2_train_packed
model | model1
dependent_varname | {y}
independent_varname | {feature_vector}
model_arch_table | model_arch_library
model_id | 1
compile_params | loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']
fit_params | batch_size=64, epochs=1
num_iterations | 10
validation_table | balanced2_test_packed
object_table |
metrics_compute_frequency | -3
name | Frank
description | Network test run
model_type | madlib_keras
model_size | 5.9853515625
start_training_time | 2021-03-12 20:24:27.74585
end_training_time | 2021-03-12 20:24:30.012898
metrics_elapsed_time | {1.13839101791382,1.57564496994019,2.02039790153503,2.26697182655334}
madlib_version | 1.18.0-dev
num_classes | {23}
dependent_vartype | {text}
normalizing_const | 1
metrics_type | {accuracy}
loss_type | categorical_crossentropy
training_metrics_final | 0.529687523841858
training_loss_final | 470.371795654297
training_metrics | {0.283894240856171,0.344591349363327,0.489182680845261,0.529687523841858}
training_loss | {3195.37939453125,1194.63610839844,508.576507568359,470.371795654297}
validation_metrics_final | 0.52836537361145
validation_loss_final | 11892.33203125
validation_metrics | {0.289903849363327,0.35432693362236,0.482692301273346,0.52836537361145}
validation_loss | {35025.0390625,23250.08984375,12720.3359375,11892.33203125}
metrics_iters | {3,6,9,10}
y_class_values | {class01,class02,class03,class04,class05,class06,class07,class08,class09,class10,class11,class12,class13,class14,class15,class16,class17,class18,class19,class20,class21,class22,normal}
{code}
Also if you make it 0 you get this cryptic error
{code}
InternalError: (psycopg2.errors.InternalError_) ZeroDivisionError: integer division or modulo by zero (plpython.c:5038)
CONTEXT: Traceback (most recent call last):
PL/Python function "madlib_keras_fit", line 23, in <module>
madlib_keras.fit(**globals())
PL/Python function "madlib_keras_fit", line 42, in wrapper
PL/Python function "madlib_keras_fit", line 273, in fit
PL/Python function "madlib_keras_fit", line 542, in should_compute_metrics_this_iter
PL/Python function "madlib_keras_fit"
[SQL: SELECT madlib.madlib_keras_fit('balanced2_train_packed', -- source table
'model1', -- model output table
'model_arch_library', -- model arch table
1, -- model arch id
$$ loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'] $$, -- compile_params
$$ batch_size=64, epochs=1 $$, -- fit_params
10, -- num_iterations
FALSE, -- use GPUs
'balanced2_test_packed', -- validation dataset
0, -- metrics compute frequency
FALSE, -- warm start
'Frank', -- name
'Network test run' -- description
);]
{code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)