Google Git
Sign in
apache / madlib-site / 0deaaf8466842a41419873ef79fe2bfef389fcab / . / docs / v1.19.0 / group__grp__keras__setup__model__selection.html
blob: 46fbced355a705f9f10c1be61d8b214cc9139081 [file] [log] [blame]
<!-- HTML header for doxygen 1.8.4-->
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/>
<meta http-equiv="X-UA-Compatible" content="IE=9"/>
<meta name="generator" content="Doxygen 1.8.13"/>
<meta name="keywords" content="madlib,postgres,greenplum,machine learning,data mining,deep learning,ensemble methods,data science,market basket analysis,affinity analysis,pca,lda,regression,elastic net,huber white,proportional hazards,k-means,latent dirichlet allocation,bayes,support vector machines,svm"/>
<title>MADlib: Define Model Configurations</title>
<link href="tabs.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="jquery.js"></script>
<script type="text/javascript" src="dynsections.js"></script>
<link href="navtree.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="resize.js"></script>
<script type="text/javascript" src="navtreedata.js"></script>
<script type="text/javascript" src="navtree.js"></script>
<script type="text/javascript">
$(document).ready(initResizable);
</script>
<link href="search/search.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="search/searchdata.js"></script>
<script type="text/javascript" src="search/search.js"></script>
<script type="text/javascript">
$(document).ready(function() { init_search(); });
</script>
<script type="text/x-mathjax-config">
MathJax.Hub.Config({
extensions: ["tex2jax.js", "TeX/AMSmath.js", "TeX/AMSsymbols.js"],
jax: ["input/TeX","output/HTML-CSS"],
});
</script><script type="text/javascript" src="http://cdn.mathjax.org/mathjax/latest/MathJax.js"></script>
<!-- hack in the navigation tree -->
<script type="text/javascript" src="eigen_navtree_hacks.js"></script>
<link href="doxygen.css" rel="stylesheet" type="text/css" />
<link href="madlib_extra.css" rel="stylesheet" type="text/css"/>
<!-- google analytics -->
<script>
(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
})(window,document,'script','//www.google-analytics.com/analytics.js','ga');
ga('create', 'UA-45382226-1', 'madlib.apache.org');
ga('send', 'pageview');
</script>
</head>
<body>
<div id="top"><!-- do not remove this div, it is closed by doxygen! -->
<div id="titlearea">
<table cellspacing="0" cellpadding="0">
<tbody>
<tr style="height: 56px;">
<td id="projectlogo"><a href="http://madlib.apache.org"><img alt="Logo" src="madlib.png" height="50" style="padding-left:0.5em;" border="0"/ ></a></td>
<td style="padding-left: 0.5em;">
<div id="projectname">
<span id="projectnumber">1.19.0</span>
</div>
<div id="projectbrief">User Documentation for Apache MADlib</div>
</td>
<td> <div id="MSearchBox" class="MSearchBoxInactive">
<span class="left">
<img id="MSearchSelect" src="search/mag_sel.png"
onmouseover="return searchBox.OnSearchSelectShow()"
onmouseout="return searchBox.OnSearchSelectHide()"
alt=""/>
<input type="text" id="MSearchField" value="Search" accesskey="S"
onfocus="searchBox.OnSearchFieldFocus(true)"
onblur="searchBox.OnSearchFieldFocus(false)"
onkeyup="searchBox.OnSearchFieldChange(event)"/>
</span><span class="right">
<a id="MSearchClose" href="javascript:searchBox.CloseResultsWindow()"><img id="MSearchCloseImg" border="0" src="search/close.png" alt=""/></a>
</span>
</div>
</td>
</tr>
</tbody>
</table>
</div>
<!-- end header part -->
<!-- Generated by Doxygen 1.8.13 -->
<script type="text/javascript">
var searchBox = new SearchBox("searchBox", "search",false,'Search');
</script>
</div><!-- top -->
<div id="side-nav" class="ui-resizable side-nav-resizable">
<div id="nav-tree">
<div id="nav-tree-contents">
<div id="nav-sync" class="sync"></div>
</div>
</div>
<div id="splitbar" style="-moz-user-select:none;"
class="ui-resizable-handle">
</div>
</div>
<script type="text/javascript">
$(document).ready(function(){initNavTree('group__grp__keras__setup__model__selection.html','');});
</script>
<div id="doc-content">
<!-- window showing the filter options -->
<div id="MSearchSelectWindow"
onmouseover="return searchBox.OnSearchSelectShow()"
onmouseout="return searchBox.OnSearchSelectHide()"
onkeydown="return searchBox.OnSearchSelectKey(event)">
</div>
<!-- iframe showing the search results (closed by default) -->
<div id="MSearchResultsWindow">
<iframe src="javascript:void(0)" frameborder="0"
name="MSearchResults" id="MSearchResults">
</iframe>
</div>
<div class="header">
<div class="headertitle">
<div class="title">Define Model Configurations<div class="ingroups"><a class="el" href="group__grp__dl.html">Deep Learning</a> &raquo; <a class="el" href="group__grp__model__selection.html">Train Multiple Models</a></div></div> </div>
</div><!--header-->
<div class="contents">
<div class="toc"><b>Contents</b><ul>
<li class="level1">
<a href="#gen_mst_configs">Generate Model Configurations</a> </li>
<li class="level1">
<a href="#load_mst_table">Load Model Selection Table [Deprecated]</a> </li>
<li class="level1">
<a href="#example">Examples</a> </li>
<li class="level1">
<a href="#related">Related Topics</a> </li>
</ul>
</div><p>This module generates model configurations using grid search or random search.</p>
<p>Once the configurations are defined, they can be used by the fit function in <a href="group__grp__keras__run__model__selection.html">Train Model Configurations</a>. By model configurations we mean both hyperparameters and model architectures. The output table from this module defines the combinations of model architectures, compile and fit parameters to be trained in parallel.</p>
<p><a class="anchor" id="gen_mst_configs"></a></p><dl class="section user"><dt>Generate Model Configurations</dt><dd></dd></dl>
<pre class="syntax">
generate_model_configs(
model_arch_table,
model_selection_table,
model_id_list,
compile_params_grid,
fit_params_grid,
search_type,
num_configs,
random_state,
object_table
)
</pre><p><b>Arguments</b> </p><dl class="arglist">
<dt>model_arch_table </dt>
<dd><p class="startdd">VARCHAR. Table containing model architectures and weights. For more information on this table refer to the module <a href="group__grp__keras__model__arch.html">Define Model Architectures</a>. </p>
<p class="enddd"></p>
</dd>
<dt>model_selection_table </dt>
<dd><p class="startdd">VARCHAR. Model selection table created by this module. If this table already exists, it will be appended to. A summary table named &lt;model_selection_table&gt;_summary is also created. Contents of both output tables are described below. </p>
<p class="enddd"></p>
</dd>
<dt>model_id_list </dt>
<dd><p class="startdd">INTEGER[]. Array of model IDs from the 'model_arch_table' to be included in the run combinations. For hyperparameter search, this will typically be one model ID. For model architecture search, this will be the different model IDs that you want to compare. </p>
<p class="enddd"></p>
</dd>
<dt>compile_params_grid </dt>
<dd><p class="startdd">VARCHAR. String representation of a Python dictionary of compile parameters to be tested. Each entry of the dictionary should consist of keys as compile parameter names, and values as a Python list of compile parameter values to be passed to Keras. Also, optimizer parameters are a nested dictionary to allow different optimizer types to have different parameters or ranges of parameters. Here is an example:</p>
<pre class="example">
$$
{'loss': ['categorical_crossentropy'],
'optimizer_params_list': [
{'optimizer': ['SGD'], 'lr': [0.0001, 0.001, 'log'], 'momentum': [0.95, 0.99, 'log_near_one']},
{'optimizer': ['Adam'], 'lr': [0.01, 0.1, 'log'], 'decay': [1e-6, 1e-4, 'log']}],
'metrics': ['accuracy']
}
$$
</pre><p>The following types of sampling are supported: 'linear', 'log' and 'log_near_one'. The 'log_near_one' sampling is useful for exponentially weighted average types of parameters like momentum, which are very sensitive to changes near 1. It has the effect of producing more values near 1 than regular log-based sampling.</p>
<p>In the case of grid search, omit the sampling type and just put the grid points in the list.</p>
<dl class="section note"><dt>Note</dt><dd><ul>
<li>Custom loss functions and custom metrics can be used as defined in <a href="group__grp__custom__function.html">Define Custom Functions.</a> List the custom function name and provide the name of the table where the serialized Python objects reside using the parameter 'object_table' below.</li>
<li>The following loss function is not supported: <em>sparse_categorical_crossentropy</em>. The following metrics are not supported: <em>sparse_categorical_accuracy, sparse_top_k_categorical_accuracy</em>.</li>
<li>The Keras accuracy parameter <em>top_k_categorical_accuracy</em> returns top 5 accuracy by default. If you want a different top k value, use the helper function <a href="group__grp__custom__function.html#top_k_function">Top k Accuracy Function</a> to create a custom Python function to compute the top k accuracy that you want. </li>
</ul>
</dd></dl>
</dd>
<dt>fit_params_grid </dt>
<dd><p class="startdd">VARCHAR. String representation of a Python dictionary of fit parameters to be tested. Each entry of the dictionary should consist of keys as fit parameter names, and values as a Python list of fit parameter values to be passed to Keras. Here is an example:</p>
<pre class="example">
$$
{'batch_size': [32, 64, 128, 256],
'epochs': [10, 20, 30]
}
$$
</pre> <dl class="section note"><dt>Note</dt><dd>Callbacks are not currently supported except for TensorBoard which you can specify in the usual way, e.g., 'callbacks': ['[TensorBoard(log_dir="/tmp/logs/fit")]'] </dd></dl>
</dd>
<dt>search_type </dt>
<dd><p class="startdd">VARCHAR, 'grid' or 'random'. Search strategy for generating model configurations. For grid search, will generate all combinations of model IDs + compile params + fit params. For random search, specify the number of configs you want to generate using the 'num_configs' parameter below. Note that you can also use short forms for the 'grid' or 'random' keywords, e.g.,'rand' or 'r' instead of writing out 'random' in full. </p>
<p class="enddd"></p>
</dd>
<dt>num_configs (optional) </dt>
<dd><p class="startdd">INTEGER, default: NULL. Number of model configs to generate. Only applies when search_type='random'. </p>
<p class="enddd"></p>
</dd>
<dt>random_state (optional) </dt>
<dd><p class="startdd">INTEGER, default: NULL. Pseudo random number generator state used for random uniform sampling from lists of possible values. Pass an integer for reproducible output across multiple function calls. Only applies when search_type='random'. </p>
<p class="enddd"></p>
</dd>
<dt>object_table (optional) </dt>
<dd><p class="startdd">VARCHAR, default: NULL. Name of the table containing Python objects in the case that custom loss functions or custom metrics are specified in the 'compile_params_grid'. Note that this table has to be created by the <a href="group__grp__custom__function.html">Define Custom Functions</a> method. It is not allowed to pass a schema name, since it will be automatically pulled from this functions associated madlib schema. </p>
<p class="enddd"></p>
</dd>
</dl>
<p><b>Output table</b> <br />
The model selection output table contains the following columns: </p><table class="output">
<tr>
<th>mst_key </th><td>INTEGER. ID that defines a unique tuple for model architecture-compile parameters-fit parameters. </td></tr>
<tr>
<th>model_id </th><td>VARCHAR. Model architecture ID from the 'model_arch_table'. </td></tr>
<tr>
<th>compile_params </th><td>VARCHAR. Keras compile parameters. </td></tr>
<tr>
<th>fit_params </th><td>VARCHAR. Keras fit parameters. </td></tr>
</table>
<p>A summary table named &lt;model_selection_table&gt;_summary is also created, which contains the following column: </p><table class="output">
<tr>
<th>model_arch_table </th><td>VARCHAR. Name of the model architecture table containing the model architecture IDs. </td></tr>
<tr>
<th>object_table </th><td>VARCHAR. Name of the object table containing the serialized Python objects for custom loss functions and custom metrics. If there are none, this field will be blank. </td></tr>
</table>
<p><a class="anchor" id="load_mst_table"></a></p><dl class="section user"><dt>Load Model Selection Table [Deprecated]</dt><dd></dd></dl>
<p>This method is deprecated and replaced by the method 'generate_model_configs' described above.</p>
<pre class="syntax">
load_model_selection_table(
model_arch_table,
model_selection_table,
model_id_list,
compile_params_list,
fit_params_list,
object_table
)
</pre><p><b>Arguments</b> </p><dl class="arglist">
<dt>model_arch_table </dt>
<dd><p class="startdd">VARCHAR. Table containing model architectures and weights. For more information on this table refer to <a href="group__grp__keras__model__arch.html">Define Model Architectures</a>. </p>
<p class="enddd"></p>
</dd>
<dt>model_selection_table </dt>
<dd><p class="startdd">VARCHAR. Model selection table created by this utility. A summary table named &lt;model_selection_table&gt;_summary is also created. Contents of both output tables are the same as described above for the method 'generate_model_configs'. </p>
<p class="enddd"></p>
</dd>
<dt>model_id_list </dt>
<dd><p class="startdd">INTEGER[]. Array of model IDs from the 'model_arch_table' to be included in the run combinations. For hyperparameter search, this will typically be one model ID. For model architecture search, this will be the different model IDs that you want to test. </p>
<p class="enddd"></p>
</dd>
<dt>compile_params_list </dt>
<dd><p class="startdd">VARCHAR[]. Array of compile parameters to be tested. Each element of the array should consist of a string of compile parameters exactly as it is to be passed to Keras. For custom loss functions or custom metrics, list the custom function name in the usual way, and also provide the name of the table where the serialized objects reside in the parameter 'object_table' below. </p>
<p class="enddd"></p>
</dd>
<dt>fit_params_list </dt>
<dd><p class="startdd">VARCHAR[]. Array of fit parameters to be tested. Each element of the array should consist of a string of fit parameters exactly as it is to be passed to Keras. Callbacks are not currently supported except for TensorBoard which you can specify in the usual way, e.g., callbacks=[TensorBoard(log_dir="/tmp/logs/fit")] </p>
<p class="enddd"></p>
</dd>
<dt>object_table (optional) </dt>
<dd><p class="startdd">VARCHAR, default: NULL. Name of the table containing Python objects in the case that custom loss functions or custom metrics are specified in the parameter 'compile_params_list'. </p>
<p class="enddd"></p>
</dd>
</dl>
<p><a class="anchor" id="example"></a></p><dl class="section user"><dt>Examples</dt><dd><ol type="1">
<li>The model selection table works in conjunction with a model architecture table, so we first create a model architecture table with two different models. Use Keras to define a model architecture with 1 hidden layer: <pre class="example">
from tensorflow import keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
model1 = Sequential()
model1.add(Dense(10, activation='relu', input_shape=(4,)))
model1.add(Dense(10, activation='relu'))
model1.add(Dense(3, activation='softmax'))
model1.summary()
<pre class="fragment">_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_1 (Dense) (None, 10) 50
_________________________________________________________________
dense_2 (Dense) (None, 10) 110
_________________________________________________________________
dense_3 (Dense) (None, 3) 33
=================================================================
Total params: 193
Trainable params: 193
Non-trainable params: 0
</pre>
</pre> Export the model to JSON: <pre class="example">
model1.to_json()
</pre> <pre class="result">
'{"class_name": "Sequential", "keras_version": "2.1.6", "config": [{"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_1", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "dtype": "float32", "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "batch_input_shape": [null, 4], "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_2", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_3", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "softmax", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 3, "use_bias": true, "activity_regularizer": null}}], "backend": "tensorflow"}'
</pre> Now use Keras to define a model architecture with 2 hidden layers: <pre class="example">
model2 = Sequential()
model2.add(Dense(10, activation='relu', input_shape=(4,)))
model2.add(Dense(10, activation='relu'))
model2.add(Dense(10, activation='relu'))
model2.add(Dense(3, activation='softmax'))
model2.summary()
<pre class="fragment">_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_4 (Dense) (None, 10) 50
_________________________________________________________________
dense_5 (Dense) (None, 10) 110
_________________________________________________________________
dense_6 (Dense) (None, 10) 110
_________________________________________________________________
dense_7 (Dense) (None, 3) 33
=================================================================
Total params: 303
Trainable params: 303
Non-trainable params: 0
</pre>
</pre> Export the model to JSON: <pre class="example">
model2.to_json()
</pre> <pre class="result">
'{"class_name": "Sequential", "keras_version": "2.1.6", "config": [{"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_4", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "dtype": "float32", "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "batch_input_shape": [null, 4], "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_5", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_6", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_7", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "softmax", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 3, "use_bias": true, "activity_regularizer": null}}], "backend": "tensorflow"}'
</pre></li>
<li>Load both models into the architecture table: <pre class="example">
DROP TABLE IF EXISTS model_arch_library;
SELECT madlib.load_keras_model('model_arch_library', -- Output table,
$$
{"class_name": "Sequential", "keras_version": "2.1.6", "config": [{"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_1", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "dtype": "float32", "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "batch_input_shape": [null, 4], "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_2", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_3", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "softmax", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 3, "use_bias": true, "activity_regularizer": null}}], "backend": "tensorflow"}
$$
::json, -- JSON blob
NULL, -- Weights
'Sophie', -- Name
'MLP with 1 hidden layer' -- Descr
);
SELECT madlib.load_keras_model('model_arch_library', -- Output table,
$$
{"class_name": "Sequential", "keras_version": "2.1.6", "config": [{"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_4", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "dtype": "float32", "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "batch_input_shape": [null, 4], "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_5", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_6", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_7", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "softmax", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 3, "use_bias": true, "activity_regularizer": null}}], "backend": "tensorflow"}
$$
::json, -- JSON blob
NULL, -- Weights
'Maria', -- Name
'MLP with 2 hidden layers' -- Descr
);
SELECT model_id, name, description FROM model_arch_library ORDER BY model_id;
</pre> <pre class="result">
model_id | name | description
----------+--------+--------------------------
1 | Sophie | MLP with 1 hidden layer
2 | Maria | MLP with 2 hidden layers
(2 rows)
</pre></li>
<li>Generate model configurations using grid search. The output table for grid search contains the unique combinations of model architectures, compile and fit parameters. <pre class="example">
DROP TABLE IF EXISTS mst_table, mst_table_summary;
SELECT madlib.generate_model_configs(
'model_arch_library', -- model architecture table
'mst_table', -- model selection table output
ARRAY[1,2], -- model ids from model architecture table
$$
{'loss': ['categorical_crossentropy'],
'optimizer_params_list': [ {'optimizer': ['Adam', 'SGD'], 'lr': [0.001, 0.01]} ],
'metrics': ['accuracy']}
$$, -- compile_param_grid
$$
{ 'batch_size': [64, 128],
'epochs': [10]
}
$$, -- fit_param_grid
'grid' -- search_type
);
SELECT * FROM mst_table ORDER BY mst_key;
</pre> <pre class="result">
mst_key | model_id | compile_params | fit_params
---------+----------+---------------------------------------------------------------------------------+--------------------------
1 | 1 | optimizer='Adam(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
2 | 1 | optimizer='Adam(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
3 | 1 | optimizer='SGD(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
4 | 1 | optimizer='SGD(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
5 | 1 | optimizer='Adam(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
6 | 1 | optimizer='Adam(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
7 | 1 | optimizer='SGD(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
8 | 1 | optimizer='SGD(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
9 | 2 | optimizer='Adam(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
10 | 2 | optimizer='Adam(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
11 | 2 | optimizer='SGD(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
12 | 2 | optimizer='SGD(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
13 | 2 | optimizer='Adam(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
14 | 2 | optimizer='Adam(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
15 | 2 | optimizer='SGD(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
16 | 2 | optimizer='SGD(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
(16 rows)
</pre> Note that above uses the same learning rate for the two optimizers. If you wanted to use different learning rates and different parameters for different optimizers (common): <pre class="example">
DROP TABLE IF EXISTS mst_table, mst_table_summary;
SELECT madlib.generate_model_configs(
'model_arch_library', -- model architecture table
'mst_table', -- model selection table output
ARRAY[1,2], -- model ids from model architecture table
$$
{'loss': ['categorical_crossentropy'],
'optimizer_params_list': [
{'optimizer': ['SGD']},
{'optimizer': ['SGD'], 'lr': [0.0001, 0.001], 'momentum': [0.95]},
{'optimizer': ['Adam'], 'lr': [0.01, 0.1], 'decay': [1e-4]}],
'metrics': ['accuracy']}
$$, -- compile_param_grid
$$
{ 'batch_size': [64, 128],
'epochs': [10]
}
$$, -- fit_param_grid
'grid' -- search_type
);
SELECT * FROM mst_table ORDER BY mst_key;
</pre> <pre class="result">
mst_key | model_id | compile_params | fit_params
---------+----------+-----------------------------------------------------------------------------------------------+--------------------------
1 | 1 | optimizer='SGD()',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
2 | 1 | optimizer='SGD()',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
3 | 2 | optimizer='SGD()',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
4 | 2 | optimizer='SGD()',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
5 | 1 | optimizer='SGD(lr=0.0001,momentum=0.95)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
6 | 1 | optimizer='SGD(lr=0.0001,momentum=0.95)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
7 | 1 | optimizer='SGD(lr=0.001,momentum=0.95)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
8 | 1 | optimizer='SGD(lr=0.001,momentum=0.95)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
9 | 2 | optimizer='SGD(lr=0.0001,momentum=0.95)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
10 | 2 | optimizer='SGD(lr=0.0001,momentum=0.95)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
11 | 2 | optimizer='SGD(lr=0.001,momentum=0.95)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
12 | 2 | optimizer='SGD(lr=0.001,momentum=0.95)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
13 | 1 | optimizer='Adam(lr=0.01,decay=0.0001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
14 | 1 | optimizer='Adam(lr=0.01,decay=0.0001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
15 | 1 | optimizer='Adam(lr=0.1,decay=0.0001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
16 | 1 | optimizer='Adam(lr=0.1,decay=0.0001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
17 | 2 | optimizer='Adam(lr=0.01,decay=0.0001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
18 | 2 | optimizer='Adam(lr=0.01,decay=0.0001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
19 | 2 | optimizer='Adam(lr=0.1,decay=0.0001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
20 | 2 | optimizer='Adam(lr=0.1,decay=0.0001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
(20 rows)
</pre></li>
<li>Generate model configurations using random search. The output table for random search contains the specified number of model architectures, compile and fit parameters, sampled from the specified distributions. <pre class="example">
DROP TABLE IF EXISTS mst_table, mst_table_summary;
SELECT madlib.generate_model_configs(
'model_arch_library', -- model architecture table
'mst_table', -- model selection table output
ARRAY[1,2], -- model ids from model architecture table
$$
{'loss': ['categorical_crossentropy'],
'optimizer_params_list': [
{'optimizer': ['SGD'], 'lr': [0.0001, 0.001, 'log'], 'momentum': [0.95, 0.99, 'log_near_one']},
{'optimizer': ['Adam'], 'lr': [0.01, 0.1, 'log'], 'decay': [1e-6, 1e-4, 'log']}],
'metrics': ['accuracy']}
$$, -- compile_param_grid
$$
{ 'batch_size': [64, 128],
'epochs': [10]
}
$$, -- fit_param_grid
'random', -- search_type
20
);
SELECT * FROM mst_table ORDER BY mst_key;
</pre> <pre class="result">
mst_key | model_id | compile_params | fit_params
---------+----------+-----------------------------------------------------------------------------------------------------------------------------+--------------------------
1 | 1 | optimizer='SGD(lr=0.000195784477708685,momentum=0.9768159513291526)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
2 | 2 | optimizer='SGD(lr=0.0002499200066875511,momentum=0.9807877269510826)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
3 | 1 | optimizer='SGD(lr=0.0009097798285407916,momentum=0.9706029152411938)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
4 | 1 | optimizer='SGD(lr=0.0001272842475986666,momentum=0.9858583458057799)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
5 | 1 | optimizer='SGD(lr=0.0001367874444015989,momentum=0.9772674033475668)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
6 | 2 | optimizer='SGD(lr=0.0002233708561319785,momentum=0.9743315606145182)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
7 | 1 | optimizer='SGD(lr=0.0009066689970530365,momentum=0.9835897505288803)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
8 | 1 | optimizer='SGD(lr=0.0007589416356572876,momentum=0.958751411608181)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
9 | 2 | optimizer='Adam(lr=0.057814228170084386,decay=1.0641718595377929e-06)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
10 | 1 | optimizer='Adam(lr=0.01927466297833838,decay=1.039476442716842e-06)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
11 | 2 | optimizer='Adam(lr=0.014718555287257804,decay=9.947768661882175e-05)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
12 | 1 | optimizer='Adam(lr=0.010397686133595378,decay=2.5730580994358942e-05)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
13 | 1 | optimizer='SGD(lr=0.0008624562426613621,momentum=0.989134963527059)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
14 | 2 | optimizer='SGD(lr=0.00010555974470031461,momentum=0.980489419269402)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
15 | 2 | optimizer='Adam(lr=0.05041699703418617,decay=4.685540619995589e-05)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
16 | 1 | optimizer='Adam(lr=0.034295140601304126,decay=1.6034699865163222e-06)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
17 | 1 | optimizer='Adam(lr=0.06888969005355218,decay=1.6318109152382423e-05)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
18 | 2 | optimizer='SGD(lr=0.0008225712651952847,momentum=0.9819748008695103)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
19 | 1 | optimizer='Adam(lr=0.0819110285922332,decay=1.6912312124827899e-06)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
20 | 1 | optimizer='Adam(lr=0.011688026325555774,decay=2.9315437856404027e-05)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
(20 rows)
</pre></li>
<li>Incremental loading for more complex combinations. If it is easier to generate the model configurations incrementally rather than all at once, you can do that by not dropping the model selection table and associated summary table, in which case the new model configurations will be appended to the existing table. Here we combine two of the previous examples in to a single output table: <pre class="example">
DROP TABLE IF EXISTS mst_table, mst_table_summary;
SELECT madlib.generate_model_configs(
'model_arch_library', -- model architecture table
'mst_table', -- model selection table output
ARRAY[1,2], -- model ids from model architecture table
$$
{'loss': ['categorical_crossentropy'],
'optimizer_params_list': [ {'optimizer': ['Adam', 'SGD'], 'lr': [0.001, 0.01]} ],
'metrics': ['accuracy']}
$$, -- compile_param_grid
$$
{ 'batch_size': [64, 128],
'epochs': [10]
}
$$, -- fit_param_grid
'grid' -- search_type
);
</pre> Now add to the existing table and note that mst_key continues where it left off: <pre class="example">
SELECT madlib.generate_model_configs(
'model_arch_library', -- model architecture table
'mst_table', -- model selection table output
ARRAY[1,2], -- model ids from model architecture table
$$
{'loss': ['categorical_crossentropy'],
'optimizer_params_list': [
{'optimizer': ['SGD'], 'lr': [0.0001, 0.001, 'log'], 'momentum': [0.95, 0.99, 'log_near_one']},
{'optimizer': ['Adam'], 'lr': [0.01, 0.1, 'log'], 'decay': [1e-6, 1e-4, 'log']}],
'metrics': ['accuracy']}
$$, -- compile_param_grid
$$
{ 'batch_size': [64, 128],
'epochs': [10]
}
$$, -- fit_param_grid
'random', -- search_type
20
);
SELECT * FROM mst_table ORDER BY mst_key;
</pre> <pre class="result">
mst_key | model_id | compile_params | fit_params
---------+----------+-----------------------------------------------------------------------------------------------------------------------------+--------------------------
1 | 1 | optimizer='Adam(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
2 | 1 | optimizer='Adam(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
3 | 1 | optimizer='SGD(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
4 | 1 | optimizer='SGD(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
5 | 1 | optimizer='Adam(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
6 | 1 | optimizer='Adam(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
7 | 1 | optimizer='SGD(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
8 | 1 | optimizer='SGD(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
9 | 2 | optimizer='Adam(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
10 | 2 | optimizer='Adam(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
11 | 2 | optimizer='SGD(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
12 | 2 | optimizer='SGD(lr=0.001)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
13 | 2 | optimizer='Adam(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
14 | 2 | optimizer='Adam(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
15 | 2 | optimizer='SGD(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
16 | 2 | optimizer='SGD(lr=0.01)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
17 | 2 | optimizer='SGD(lr=0.00013996842804647915,momentum=0.9677072493281305)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
18 | 1 | optimizer='Adam(lr=0.04252873277972123,decay=9.503983307511243e-05)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
19 | 2 | optimizer='Adam(lr=0.06666969394323848,decay=1.5626668941131748e-05)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
20 | 2 | optimizer='SGD(lr=0.00016137313867804707,momentum=0.954293112127019)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
21 | 2 | optimizer='Adam(lr=0.019443570245321506,decay=1.2882524497407873e-06)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
22 | 2 | optimizer='Adam(lr=0.06302317748060839,decay=6.238009849562074e-05)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
23 | 2 | optimizer='SGD(lr=0.00010890482493011119,momentum=0.9826239169968034)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
24 | 1 | optimizer='SGD(lr=0.0009201966766121783,momentum=0.9896730563556151)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
25 | 2 | optimizer='SGD(lr=0.00028961522836420906,momentum=0.9859394149216544)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
26 | 1 | optimizer='SGD(lr=0.0001503249757866609,momentum=0.9777816636354879)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
27 | 2 | optimizer='SGD(lr=0.0008405326172626768,momentum=0.9538686498263182)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
28 | 1 | optimizer='SGD(lr=0.00011926989091387571,momentum=0.9876746918399469)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
29 | 1 | optimizer='Adam(lr=0.018794361633022855,decay=9.387826286694454e-06)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
30 | 2 | optimizer='SGD(lr=0.0009692977025027591,momentum=0.9878758592330659)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
31 | 2 | optimizer='SGD(lr=0.0006671929498585603,momentum=0.9786502962872058)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
32 | 2 | optimizer='Adam(lr=0.03948766165185474,decay=3.056584635386748e-06)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
33 | 2 | optimizer='Adam(lr=0.020343961099103417,decay=1.183810228780669e-05)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
34 | 1 | optimizer='Adam(lr=0.016854644990148417,decay=3.561117893117444e-06)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
35 | 2 | optimizer='SGD(lr=0.0004620089560788749,momentum=0.9887310587871919)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
36 | 1 | optimizer='SGD(lr=0.0002493912675066962,momentum=0.9892077270385708)',metrics=['accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
(36 rows)
</pre></li>
<li>Create model selection table manually. If you want more control over the content of the model selection table, you could use grid or random search to generate a large number of combinations, then SELECT a subset of rows for training. Alternatively, you could manually create the model selection table and the associated summary table. Both must be created since they are needed by the multiple model fit module. For example, let's say we don't want all combinations but only want batch_size=4 for model_id=1 and batch_size=8 for model_id=2: <pre class="example">
DROP TABLE IF EXISTS mst_table_manual;
CREATE TABLE mst_table_manual(
mst_key serial,
model_id integer,
compile_params varchar,
fit_params varchar
);
INSERT INTO mst_table_manual(model_id, compile_params, fit_params) VALUES
(1, $$loss='categorical_crossentropy',optimizer='Adam(lr=0.1)',metrics=['accuracy']$$, 'batch_size=4,epochs=1'),
(1, $$loss='categorical_crossentropy',optimizer='Adam(lr=0.01)',metrics=['accuracy']$$, 'batch_size=4,epochs=1'),
(1, $$loss='categorical_crossentropy',optimizer='Adam(lr=0.001)',metrics=['accuracy']$$, 'batch_size=4,epochs=1'),
(2, $$loss='categorical_crossentropy',optimizer='Adam(lr=0.1)',metrics=['accuracy']$$, 'batch_size=8,epochs=1'),
(2, $$loss='categorical_crossentropy',optimizer='Adam(lr=0.01)',metrics=['accuracy']$$, 'batch_size=8,epochs=1'),
(2, $$loss='categorical_crossentropy',optimizer='Adam(lr=0.001)',metrics=['accuracy']$$, 'batch_size=8,epochs=1');
SELECT * FROM mst_table_manual ORDER BY mst_key;
</pre> <pre class="result">
mst_key | model_id | compile_params | fit_params
---------+---------------+---------------------------------------------------------------------------------+-----------------------
1 | 1 | loss='categorical_crossentropy',optimizer='Adam(lr=0.1)',metrics=['accuracy'] | batch_size=4,epochs=1
2 | 1 | loss='categorical_crossentropy',optimizer='Adam(lr=0.01)',metrics=['accuracy'] | batch_size=4,epochs=1
3 | 1 | loss='categorical_crossentropy',optimizer='Adam(lr=0.001)',metrics=['accuracy'] | batch_size=4,epochs=1
4 | 2 | loss='categorical_crossentropy',optimizer='Adam(lr=0.1)',metrics=['accuracy'] | batch_size=8,epochs=1
5 | 2 | loss='categorical_crossentropy',optimizer='Adam(lr=0.01)',metrics=['accuracy'] | batch_size=8,epochs=1
6 | 2 | loss='categorical_crossentropy',optimizer='Adam(lr=0.001)',metrics=['accuracy'] | batch_size=8,epochs=1
(6 rows)
</pre> Create the summary table which must be named with the model selection output table appended by "_summary": <pre class="example">
DROP TABLE IF EXISTS mst_table_manual_summary;
CREATE TABLE mst_table_manual_summary (
model_arch_table varchar
);
INSERT INTO mst_table_manual_summary(model_arch_table) VALUES
('model_arch_library');
SELECT * FROM mst_table_manual_summary;
</pre> <pre class="result">
model_arch_table
--------------------+
model_arch_library
(1 row)
</pre></li>
<li>Custom loss functions and custom metrics. Let's say we have a table 'custom_function_table' that contains a custom loss function called 'my_custom_loss' and a custom accuracy function called 'my_custom_accuracy' based on <a href="group__grp__custom__function.html">Define Custom Functions.</a> Generate the model configurations with: <pre class="example">
DROP TABLE IF EXISTS mst_table, mst_table_summary;
SELECT madlib.generate_model_configs(
'model_arch_library', -- model architecture table
'mst_table', -- model selection table output
ARRAY[1,2], -- model ids from model architecture table
$$
{'loss': ['my_custom_loss'],
'optimizer_params_list': [ {'optimizer': ['Adam', 'SGD'], 'lr': [0.001, 0.01]} ],
'metrics': ['my_custom_accuracy']}
$$, -- compile_param_grid
$$
{ 'batch_size': [64, 128],
'epochs': [10]
}
$$, -- fit_param_grid
'grid', -- search_type
NULL, -- num_configs
NULL, -- random_state
'custom_function_table' -- table with custom functions
);
SELECT * FROM mst_table ORDER BY mst_key;
</pre> <pre class="result">
mst_key | model_id | compile_params | fit_params
---------+----------+---------------------------------------------------------------------------------+--------------------------
1 | 1 | optimizer='Adam(lr=0.001)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=64
2 | 1 | optimizer='Adam(lr=0.001)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=128
3 | 1 | optimizer='SGD(lr=0.001)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=64
4 | 1 | optimizer='SGD(lr=0.001)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=128
5 | 1 | optimizer='Adam(lr=0.01)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=64
6 | 1 | optimizer='Adam(lr=0.01)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=128
7 | 1 | optimizer='SGD(lr=0.01)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=64
8 | 1 | optimizer='SGD(lr=0.01)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=128
9 | 2 | optimizer='Adam(lr=0.001)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=64
10 | 2 | optimizer='Adam(lr=0.001)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=128
11 | 2 | optimizer='SGD(lr=0.001)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=64
12 | 2 | optimizer='SGD(lr=0.001)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=128
13 | 2 | optimizer='Adam(lr=0.01)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=64
14 | 2 | optimizer='Adam(lr=0.01)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=128
15 | 2 | optimizer='SGD(lr=0.01)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=64
16 | 2 | optimizer='SGD(lr=0.01)',metrics=['my_custom_accuracy'],loss='my_custom_loss' | epochs=10,batch_size=128
(16 rows)
</pre> Similarly, if you created a custom top k categorical accuracy function 'top_3_accuracy' in <a href="group__grp__custom__function.html">Define Custom Functions</a> you can generate the model configurations as: <pre class="example">
DROP TABLE IF EXISTS mst_table, mst_table_summary;
SELECT madlib.generate_model_configs(
'model_arch_library', -- model architecture table
'mst_table', -- model selection table output
ARRAY[1,2], -- model ids from model architecture table
$$
{'loss': ['categorical_crossentropy'],
'optimizer_params_list': [ {'optimizer': ['Adam', 'SGD'], 'lr': [0.001, 0.01]} ],
'metrics': ['top_3_accuracy']}
$$, -- compile_param_grid
$$
{ 'batch_size': [64, 128],
'epochs': [10]
}
$$, -- fit_param_grid
'grid', -- search_type
NULL, -- num_configs
NULL, -- random_state
'custom_function_table' -- table with custom functions
);
SELECT * FROM mst_table ORDER BY mst_key;
</pre> <pre class="result">
mst_key | model_id | compile_params | fit_params
---------+----------+---------------------------------------------------------------------------------+--------------------------
1 | 1 | optimizer='Adam(lr=0.001)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
2 | 1 | optimizer='Adam(lr=0.001)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
3 | 1 | optimizer='SGD(lr=0.001)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
4 | 1 | optimizer='SGD(lr=0.001)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
5 | 1 | optimizer='Adam(lr=0.01)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
6 | 1 | optimizer='Adam(lr=0.01)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
7 | 1 | optimizer='SGD(lr=0.01)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
8 | 1 | optimizer='SGD(lr=0.01)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
9 | 2 | optimizer='Adam(lr=0.001)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
10 | 2 | optimizer='Adam(lr=0.001)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
11 | 2 | optimizer='SGD(lr=0.001)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
12 | 2 | optimizer='SGD(lr=0.001)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
13 | 2 | optimizer='Adam(lr=0.01)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
14 | 2 | optimizer='Adam(lr=0.01)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
15 | 2 | optimizer='SGD(lr=0.01)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=64
16 | 2 | optimizer='SGD(lr=0.01)',metrics=['top_3_accuracy'],loss='categorical_crossentropy' | epochs=10,batch_size=128
(16 rows)
</pre></li>
<li><b>[Deprecated]</b> Load model selection table. This method is replaced by the 'generate_model_configs' method described above. Select the model(s) from the model architecture table that you want to run, along with the compile and fit parameters. Unique combinations will be created: <pre class="example">
DROP TABLE IF EXISTS mst_table, mst_table_summary;
SELECT madlib.load_model_selection_table('model_arch_library', -- model architecture table
'mst_table', -- model selection table output
ARRAY[1,2], -- model ids from model architecture table
ARRAY[ -- compile params
$$loss='categorical_crossentropy',optimizer='Adam(lr=0.1)',metrics=['accuracy']$$,
$$loss='categorical_crossentropy', optimizer='Adam(lr=0.01)',metrics=['accuracy']$$,
$$loss='categorical_crossentropy',optimizer='Adam(lr=0.001)',metrics=['accuracy']$$
],
ARRAY[ -- fit params
$$batch_size=4,epochs=1$$,
$$batch_size=8,epochs=1$$
]
);
SELECT * FROM mst_table ORDER BY mst_key;
</pre> <pre class="result">
mst_key | model_id | compile_params | fit_params
---------+---------------+---------------------------------------------------------------------------------+-----------------------
1 | 1 | loss='categorical_crossentropy',optimizer='Adam(lr=0.1)',metrics=['accuracy'] | batch_size=4,epochs=1
2 | 1 | loss='categorical_crossentropy',optimizer='Adam(lr=0.1)',metrics=['accuracy'] | batch_size=8,epochs=1
3 | 1 | loss='categorical_crossentropy', optimizer='Adam(lr=0.01)',metrics=['accuracy'] | batch_size=4,epochs=1
4 | 1 | loss='categorical_crossentropy', optimizer='Adam(lr=0.01)',metrics=['accuracy'] | batch_size=8,epochs=1
5 | 1 | loss='categorical_crossentropy',optimizer='Adam(lr=0.001)',metrics=['accuracy'] | batch_size=4,epochs=1
6 | 1 | loss='categorical_crossentropy',optimizer='Adam(lr=0.001)',metrics=['accuracy'] | batch_size=8,epochs=1
7 | 2 | loss='categorical_crossentropy',optimizer='Adam(lr=0.1)',metrics=['accuracy'] | batch_size=4,epochs=1
8 | 2 | loss='categorical_crossentropy',optimizer='Adam(lr=0.1)',metrics=['accuracy'] | batch_size=8,epochs=1
9 | 2 | loss='categorical_crossentropy', optimizer='Adam(lr=0.01)',metrics=['accuracy'] | batch_size=4,epochs=1
10 | 2 | loss='categorical_crossentropy', optimizer='Adam(lr=0.01)',metrics=['accuracy'] | batch_size=8,epochs=1
11 | 2 | loss='categorical_crossentropy',optimizer='Adam(lr=0.001)',metrics=['accuracy'] | batch_size=4,epochs=1
12 | 2 | loss='categorical_crossentropy',optimizer='Adam(lr=0.001)',metrics=['accuracy'] | batch_size=8,epochs=1
(12 rows)
</pre> The name of the model architecture table is stored in the summary table: <pre class="example">
SELECT * FROM mst_table_summary;
</pre> <pre class="result">
model_arch_table
--------------------+
model_arch_library
(1 row)
</pre></li>
</ol>
</dd></dl>
<p><a class="anchor" id="related"></a></p><dl class="section user"><dt>Related Topics</dt><dd></dd></dl>
<p>See <a class="el" href="keras__model__arch__table_8sql__in.html">keras_model_arch_table.sql_in</a> </p>
</div><!-- contents -->
</div><!-- doc-content -->
<!-- start footer part -->
<div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
<ul>
<li class="footer">Generated on Wed Dec 15 2021 20:27:20 for MADlib by
<a href="http://www.doxygen.org/index.html">
<img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.13 </li>
</ul>
</div>
</body>
</html>