/* ----------------------------------------------------------------------- *//**
 *
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you under the Apache License, Version 2.0 (the
 * "License"); you may not use this file except in compliance
 * with the License.  You may obtain a copy of the License at
 *
 *   http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
 * KIND, either express or implied.  See the License for the
 * specific language governing permissions and limitations
 * under the License.
 *
 *
 * @file madlib_keras.sql_in
 *
 * @brief SQL functions for distributed deep learning with keras
 * @date June 2019
 *
 *
 *//* ----------------------------------------------------------------------- */

m4_include(`SQLCommon.m4')

/**
@addtogroup grp_keras

@brief Fit, evaluate and predict using the Keras API.

<div class="toc"><b>Contents</b><ul>
<li class="level1"><a href="#keras_fit">Fit</a></li>
<li class="level1"><a href="#keras_evaluate">Evaluate</a></li>
<li class="level1"><a href="#keras_predict">Predict</a></li>
<li class="level1"><a href="#keras_predict_byom">Predict BYOM</a></li>
<li class="level1"><a href="#example">Examples</a></li>
<li class="level1"><a href="#notes">Notes</a></li>
<li class="level1"><a href="#background">Technical Background</a></li>
<li class="level1"><a href="#literature">Literature</a></li>
<li class="level1"><a href="#related">Related Topics</a></li>
</ul></div>

\warning <em> This MADlib method is still in early stage development.
Interface and implementation are subject to change. </em>

This module allows you to use SQL to call deep learning
models designed in Keras [1], which is a high-level neural
network API written in Python.
Keras was developed for fast experimentation.  It can run
on top of different backends and the one that is currently
supported by MADlib is TensorFlow [2].  The implementation
in MADlib is distributed and designed to train
a single model across multiple segments (workers)
in Greenplum database.  (PostgreSQL is also supported.)
Alternatively, to train multiple models at the same time for model
architecture search or hyperparameter tuning, you can
use <a href="group__grp__keras__run__model__selection.html">Model Selection</a>.

The main use case is image classification
using sequential models, which are made up of a
linear stack of layers.  This includes multilayer perceptrons (MLPs)
and convolutional neural networks (CNNs).  Regression is not
currently supported.

Before using Keras in MADlib you will need to mini-batch
your training and evaluation datasets by calling the
<a href="group__grp__input__preprocessor__dl.html">Preprocessor
for Images</a> which is a utility that prepares image data for
use by models that support mini-batch as an optimization option.
This is a one-time operation and you would only
need to re-run the preprocessor if your input data has changed.
The advantage of using mini-batching is that it
can perform better than stochastic gradient descent
because it uses more than one training example at a time,
typically resulting faster and smoother convergence [3].

You can also do inference on models that have not been trained in MADlib,
but rather imported from an external source.  This is in the section
called "Predict BYOM" below, where "BYOM" stands for "Bring Your Own Model."

Note that the following MADlib functions are targeting a specific Keras
version (2.2.4) with a specific TensorFlow kernel version (1.14).
Using a newer or older version may or may not work as intended.

@note CUDA GPU memory cannot be released until the process holding it is terminated. 
When a MADlib deep learning function is called with GPUs, Greenplum internally 
creates a process (called a slice) which calls TensorFlow to do the computation. 
This process holds the GPU memory until one of the following two things happen:
query finishes and user logs out of the Postgres client/session; or, 
query finishes and user waits for the timeout set by `gp_vmem_idle_resource_timeout`.  
The default value for this timeout is 18 sec [8].  So the recommendation is:
log out/reconnect to the session after every GPU query; or
wait for `gp_vmem_idle_resource_timeout` before you run another GPU query (you can 
also set it to a lower value).

@anchor keras_fit
@par Fit
The fit (training) function has the following format:

<pre class="syntax">
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
    )
</pre>

\b Arguments
<dl class="arglist">
  <dt>source_table</dt>
  <dd>TEXT. Name of the table containing the training data.
  This is the name of the output
  table from the image preprocessor.  Independent
  and dependent variables are specified in the preprocessor
  step which is why you do not need to explictly state
  them here as part of the fit function.</dd>

  <dt>model</dt>
  <dd>TEXT. Name of the output table containing the model.
  Details of the output table are shown below.
  </dd>

  <dt>model_arch_table</dt>
  <dd>TEXT. Name of the table containing the model
  architecture and (optionally) initial weights to use for
  training.
  </dd>

  <dt>model_id</dt>
  <dd>INTEGER. This is the id in 'model_arch_table'
  containing the model architecture and (optionally)
  initial weights to use for training.
  </dd>

  <DT>compile_params</DT>
  <DD>TEXT.
    Parameters passed to the compile method of the Keras
    model class [4].  These parameters will be passed through as is
    so they must conform to the Keras API definition.
    As an example, you might use something like: <em>loss='categorical_crossentropy', optimizer='adam', metrics=['acc']</em>.
    The mandatory parameters that must be specified are 'optimizer'
    and 'loss'.  Others are optional and will use the default
    values as per Keras if not specified here.  Also, when
    specifying 'loss' and 'metrics' do <em>not</em> include the
    module and submodule prefixes
    like <em>loss='losses.categorical_crossentropy'</em>
    or <em>optimizer='keras.optmizers.adam'</em>.

    @note
    The following loss function is
    not supported: <em>sparse_categorical_crossentropy</em>.
    The following metrics are not
    supported: <em>sparse_categorical_accuracy, top_k_categorical_accuracy, sparse_top_k_categorical_accuracy</em> and custom metrics.

  </DD>

  <DT>fit_params </DT>
  <DD>TEXT.  Parameters passed to the fit method of the Keras
    model class [4].  These will be passed through as is
    so they must conform to the Keras API definition.
    As an example, you might use something like:
    <em>batch_size=128, epochs=4</em>.
    There are no mandatory parameters so
    if you specify NULL, it will use all default
    values as per Keras.
  </DD>

  <DT>num_iterations</DT>
  <DD>INTEGER.  Number of iterations to train.
  </DD>

  <DT>use_gpus (optional)</DT>
  <DD>BOOLEAN, default: FALSE (i.e., CPU). Determines whether GPUs
  are to be used for training the neural network.  Set to TRUE to use GPUs.

    @note
    This parameter must not conflict with how the distribution rules are set in
    the preprocessor function.  For example, if you set a distribution rule to use
    certain segments on hosts that do not have GPUs attached, you will get an error
    if you set ‘use_gpus’ to TRUE.  Also, we have seen some memory related issues
    when segments share GPU resources.
    For example, if you have 1 GPU per segment host and your cluster has 4
    segments per segment host, it means that all 4
    segments will share the same
    GPU on each host. The current recommended
    configuration is 1 GPU per segment.
  </DD>

  <dt>validation_table (optional)</dt>
  <dd>TEXT, default: none. Name of the table containing
  the validation dataset.
  Note that the validation dataset must be preprocessed
  in the same way as the training dataset, so this
  is the name of the output
  table from running the image preprocessor on the validation dataset.
  Using a validation dataset can mean a
  longer training time, depending on its size.
  This can be controlled using the 'metrics_compute_frequency'
  paremeter described below.</dd>

  <DT>metrics_compute_frequency (optional)</DT>
  <DD>INTEGER, default: once at the end of training
  after 'num_iterations'.  Frequency to compute per-iteration
  metrics for the training dataset and validation dataset
  (if specified).  There can be considerable cost to
  computing metrics every iteration, especially if the
  training dataset is large.  This parameter is a way of
  controlling the frequency of those computations.
  For example, if you specify 5, then metrics will be computed
  every 5 iterations as well as at the end of training
  after 'num_iterations'.  If you use the default,
  metrics will be computed only
  once after 'num_iterations' have completed.
  </DD>

  <DT>warm_start (optional)</DT>
  <DD>BOOLEAN, default: FALSE.
    Initalize weights with the coefficients
    from the last call of the fit
    function. If set to TRUE, weights will be
    initialized from the model table
    generated by the previous training run.

    @note
    The warm start feature works based on the name of the
    model output table from a previous training run.
    When using warm start, do not drop the model output table
    or the model output summary table
    before calling the fit function, since these are needed to obtain the
    weights from the previous run.
    If you are not using warm start, the model output table
    and the model output table summary must be dropped in
    the usual way before calling the training function.
  </DD>

  <DT>name (optional)</DT>
  <DD>TEXT, default: NULL.
    Free text string to identify a name, if desired.
  </DD>

  <DT>description (optional)</DT>
  <DD>TEXT, default: NULL.
    Free text string to provide a description, if desired.
  </DD>
</dl>

<b>Output tables</b>
<br>
    The model table produced by fit contains the following columns:
    <table class="output">
      <tr>
        <th>model_weights</th>
        <td>BYTEA8. Byte array containing the weights of the neural net.</td>
      </tr>
      <tr>
        <th>model_arch</th>
        <td>TEXT. A JSON representation of the model architecture
        used in training.</td>
      </tr>
    </table>

A summary table named \<model\>_summary is also created, which has the following columns:
    <table class="output">
    <tr>
        <th>source_table</th>
        <td>Source table used for training.</td>
    </tr>
    <tr>
        <th>model</th>
        <td>Model output table produced by training.</td>
    </tr>
    <tr>
        <th>independent_varname</th>
        <td>Independent variables column from the original
        source table in the image preprocessing step.</td>
    </tr>
    <tr>
        <th>dependent_varname</th>
        <td>Dependent variable column from the original
        source table in the image preprocessing step.</td>
    </tr>
    <tr>
        <th>model_arch_table</th>
        <td>Name of the table containing
        the model architecture and (optionally) the
        initial model weights.</td>
    </tr>
    <tr>
        <th>model_id</th>
        <td>The id of the model in
        the model architecture table used for training.</td>
    </tr>
    <tr>
        <th>compile_params</th>
        <td>Compile parameters passed to Keras.</td>
    </tr>
    <tr>
        <th>fit_params</th>
        <td>Fit parameters passed to Keras.</td>
    </tr>
    <tr>
        <th>num_iterations</th>
        <td>Number of iterations of training completed.</td>
    </tr>
    <tr>
        <th>validation_table</th>
        <td>Name of the table containing
        the validation dataset (if specified).</td>
    </tr>
    <tr>
        <th>metrics_compute_frequency</th>
        <td>Frequency that per-iteration metrics are computed
        for the training dataset and validation
        dataset.</td>
    </tr>
    <tr>
        <th>name</th>
        <td>Name of the training run (free text).</td>
    </tr>
    <tr>
        <th>description</th>
        <td>Description of the training run (free text).</td>
    </tr>
    <tr>
        <th>model_type</th>
        <td>General identifier for type of model trained.
        Currently says 'madlib_keras'.</td>
    </tr>
    <tr>
        <th>model_size</th>
        <td>Size of the model in KB.  Models are stored in
        'bytea' data format which is used for binary strings
        in PostgreSQL type databases.</td>
    </tr>
    <tr>
        <th>start_training_time</th>
        <td>Timestamp for start of training.</td>
    </tr>
    <tr>
        <th>end_training_time</th>
        <td>Timestamp for end of training.</td>
    </tr>
    <tr>
        <th>metrics_elapsed_time</th>
        <td> Array of elapsed time for metric computations as
        per the 'metrics_compute_frequency' parameter.
        Useful for drawing a curve showing loss, accuracy or
        other metrics as a function of time.
        For example, if 'metrics_compute_frequency=5'
        this would be an array of elapsed time for every 5th
        iteration, plus the last iteration.</td>
    </tr>
    <tr>
        <th>madlib_version</th>
        <td>Version of MADlib used.</td>
    </tr>
    <tr>
        <th>num_classes</th>
        <td>Count of distinct classes values used.</td>
    </tr>
    <tr>
        <th>class_values</th>
        <td>Array of actual class values used.</td>
    </tr>
    <tr>
        <th>dependent_vartype</th>
        <td>Data type of the dependent variable.</td>
    </tr>
    <tr>
        <th>normalizing_constant</th>
        <td>Normalizing constant used from the
        image preprocessing step.</td>
    </tr>
    <tr>
        <th>metrics_type</th>
        <td>Metric specified in the 'compile_params'.</td>
    </tr>
    <tr>
        <th>training_metrics_final</th>
        <td>Final value of the training
        metric after all iterations have completed.
        The metric reported is the one
        specified in the 'metrics_type' parameter.</td>
    </tr>
    <tr>
        <th>training_loss_final</th>
        <td>Final value of the training loss after all
        iterations have completed.</td>
    </tr>
    <tr>
        <th>training_metrics</th>
        <td>Array of training metrics as
        per the 'metrics_compute_frequency' parameter.
        For example, if 'metrics_compute_frequency=5'
        this would be an array of metrics for every 5th
        iteration, plus the last iteration.</td>
    </tr>
    <tr>
        <th>training_loss</th>
        <td>Array of training losses as
        per the 'metrics_compute_frequency' parameter.
        For example, if 'metrics_compute_frequency=5'
        this would be an array of losses for every 5th
        iteration, plus the last iteration.</td>
    </tr>
    <tr>
        <th>validation_metrics_final</th>
        <td>Final value of the validation
        metric after all iterations have completed.
        The metric reported is the one
        specified in the 'metrics_type' parameter.</td>
    </tr>
    <tr>
        <th>validation_loss_final</th>
        <td>Final value of the validation loss after all
        iterations have completed.</td>
    </tr>
    <tr>
        <th>validation_metrics</th>
        <td>Array of validation metrics as
        per the 'metrics_compute_frequency' parameter.
        For example, if 'metrics_compute_frequency=5'
        this would be an array of metrics for every 5th
        iteration, plus the last iteration.</td>
    </tr>
    <tr>
        <th>validation_loss</th>
        <td>Array of validation losses as
        per the 'metrics_compute_frequency' parameter.
        For example, if 'metrics_compute_frequency=5'
        this would be an array of losses for every 5th
        iteration, plus the last iteration.</td>
    </tr>
    <tr>
        <th>metrics_iters</th>
        <td>Array indicating the iterations for which
        metrics are calculated, as derived from the
        parameters 'num_iterations' and 'metrics_compute_frequency'.
        For example, if 'num_iterations=5'
        and 'metrics_compute_frequency=2', then 'metrics_iters' value
        would be {2,4,5} indicating that metrics were computed
        at iterations 2, 4 and 5 (at the end).
        If 'num_iterations=5'
        and 'metrics_compute_frequency=1', then 'metrics_iters' value
        would be {1,2,3,4,5} indicating that metrics were computed
        at every iteration.</td>
    </tr>
   </table>

@anchor keras_evaluate
@par Evaluate
The evaluation function has the following format:

<pre class="syntax">
madlib_keras_evaluate(
    model_table,
    test_table,
    output_table,
    use_gpus,
    mst_key
    )
</pre>

\b Arguments
<dl class="arglist">

<DT>model_table</DT>
  <DD>TEXT. Name of the table containing the model
  to use for validation.
  </DD>

  <DT>test_table</DT>
  <dd>TEXT. Name of the table containing the evaluation dataset.
  Note that test/validation data must be preprocessed in the same
  way as the training dataset, so
  this is the name of the output
  table from the image preprocessor.  Independent
  and dependent variables are specified in the preprocessor
  step which is why you do not need to explictly state
  them here as part of the fit function.</dd>

  <DT>output_table</DT>
  <DD>TEXT. Name of table that validation output will be
  written to. Table contains:</DD>
    <table class="output">
      <tr>
        <th>loss</th>
        <td>Loss value on evaluation dataset.</td>
      </tr>
      <tr>
        <th>metric</th>
        <td>Metric value on evaluation dataset, where 'metrics_type'
        below identifies the type of metric.</td>
      </tr>
      <tr>
        <th>metrics_type</th>
        <td>Type of metric used that was used in the training step.</td>
      </tr>

  <DT>use_gpus (optional)</DT>
  <DD>BOOLEAN, default: FALSE (i.e., CPU). Determines whether GPUs
  are to be used for training the neural network.  Set to TRUE to use GPUs.

    @note
    This parameter must not conflict with how the distribution rules are set in
    the preprocessor function.  For example, if you set a distribution rule to use
    certain segments on hosts that do not have GPUs attached, you will get an error
    if you set ‘use_gpus’ to TRUE.  Also, we have seen some memory related issues
    when segments share GPU resources.
    For example, if you have 1 GPU per segment host and your cluster has 4
    segments per segment host, it means that all 4
    segments will share the same
    GPU on each host. The current recommended
    configuration is 1 GPU per segment.
  </DD>

  <DT>mst_key (optional)</DT>
  <DD>INTEGER, default: NULL. ID that defines a unique tuple for
  model architecture-compile parameters-fit parameters in a model
  selection table.  Do not use this if training one model at a time using madlib_keras_fit().
  See the <a href="group__grp__keras__run__model__selection.html">Model Selection</a> section
  for more details on model selection by training multiple models at a time.
  </DD>
</DL>

@anchor keras_predict
@par Predict
The prediction function has the following format:
<pre class="syntax">
madlib_keras_predict(
    model_table,
    test_table,
    id_col,
    independent_varname,
    output_table,
    pred_type,
    use_gpus,
    mst_key
    )
</pre>

\b Arguments
<dl class="arglist">

<DT>model_table</DT>
  <DD>TEXT. Name of the table containing the model
  to use for prediction.
  </DD>

  <DT>test_table</DT>
  <DD>TEXT. Name of the table containing the dataset to
  predict on.  Note that test data is not preprocessed (unlike
  fit and evaluate) so put one test image per row for prediction.
  Also see the comment below for the 'independent_varname' parameter
  regarding normalization.

  </DD>

  <DT>id_col</DT>
  <DD>TEXT. Name of the id column in the test data table.
  </DD>

  <DT>independent_varname</DT>
  <DD>TEXT. Column with independent variables in the test table.
  If a 'normalizing_const' is specified when preprocessing the
  training dataset, this same normalization will be applied to
  the independent variables used in predict.
  </DD>

  <DT>output_table</DT>
  <DD>TEXT. Name of the table that prediction output will be
  written to. Table contains:</DD>
    <table class="output">
      <tr>
        <th>id</th>
        <td>Gives the 'id' for each prediction, corresponding to each row from the test_table.</td>
      </tr>
      <tr>
        <th>estimated_COL_NAME</th>
        <td>
        (For pred_type='response') The estimated class
         for classification, where
         COL_NAME is the name of the column to be
         predicted from test data.
        </td>
      </tr>
      <tr>
        <th>prob_CLASS</th>
        <td>
        (For pred_type='prob' for classification) The
        probability of a given class.
        There will be one column for each class
        in the training data.
        </td>
      </tr>

  <DT>pred_type (optional)</DT>
  <DD>TEXT, default: 'response'. The type of output
  desired, where 'response' gives the actual prediction
  and 'prob' gives the probability value for each class.
  </DD>

  <DT>use_gpus (optional)</DT>
  <DD>BOOLEAN, default: FALSE (i.e., CPU).
    Flag to enable GPU support for training neural network.
    The number of GPUs to use is determined by the parameters
    passed to the preprocessor.

    @note
    We have seen some memory related issues when segments
    share GPU resources.
    For example, if you provide 1 GPU and your
    database cluster is set up to have 4
    segments per segment host, it means that all 4
    segments on a segment host will share the same
    GPU. The current recommended
    configuration is 1 GPU per segment.
  </DD>

  <DT>mst_key (optional)</DT>
  <DD>INTEGER, default: NULL. ID that defines a unique tuple for
  model architecture-compile parameters-fit parameters in a model
  selection table.  Do not use this if training one model at a time using madlib_keras_fit().
  See the <a href="group__grp__keras__run__model__selection.html">Model Selection</a> section
  for more details on model selection by training multiple models at a time.
  </DD>
</DL>


@anchor keras_predict_byom
@par Predict BYOM (bring your own model)
The predict BYOM function allows you to do inference on models that
have not been trained on MADlib, but rather imported from elsewhere.
It has the following format:
<pre class="syntax">
madlib_keras_predict_byom(
    model_arch_table,
    model_id,
    test_table,
    id_col,
    independent_varname,
    output_table,
    pred_type,
    use_gpus,
    class_values,
    normalizing_const
    )
</pre>


\b Arguments
<dl class="arglist">

<DT>model_arch_table</DT>
  <DD>TEXT. Name of the architecture table containing the model
  to use for prediction. The model weights and architecture can be loaded to
  this table by using the
  <a href="group__grp__keras__model__arch.html">load_keras_model</a> function.
  </DD>

  <DT>model_id</DT>
  <DD>INTEGER. This is the id in 'model_arch_table' containing the model
  architecture and model weights to use for prediction.
  </DD>

  <DT>test_table</DT>
  <DD>TEXT. Name of the table containing the dataset to
  predict on.  Note that test data is not preprocessed (unlike
  fit and evaluate) so put one test image per row for prediction.
  Set the 'normalizing_const' below for the independent variable if necessary.
  </DD>

  <DT>id_col</DT>
  <DD>TEXT. Name of the id column in the test data table.
  </DD>

  <DT>independent_varname</DT>
  <DD>TEXT. Column with independent variables in the test table.
  Set the 'normalizing_const' below if necessary.
  </DD>

  <DT>output_table</DT>
  <DD>TEXT. Name of the table that prediction output will be
  written to. Table contains:</DD>
    <table class="output">
      <tr>
        <th>id</th>
        <td>Gives the 'id' for each prediction, corresponding to each row from the 'test_table'.</td>
      </tr>
      <tr>
        <th>estimated_dependent_var</th>
        <td>
        (For pred_type='response') Estimated class for classification. If
        the 'class_values' parameter is passed in as NULL, then we assume that the class
        labels are [0,1,2...,n-1] where n-1 is the number of classes in the model
        architecture.
        </td>
      </tr>
      <tr>
        <th>prob_CLASS</th>
        <td>
         (For pred_type='prob' for classification)
         Probability of a given class.
         If 'class_values' is passed in as NULL, we create one column called
         'prob' which is an array of probabilities for each class.
         If 'class_values' is not NULL, then there will be one
         column for each class.
        </td>
      </tr>

  <DT>pred_type (optional)</DT>
  <DD>TEXT, default: 'response'. The type of output desired, where 'response'
   gives the actual prediction and 'prob' gives the probability value for each class.
  </DD>

  <DT>use_gpus (optional)</DT>
  <DD>BOOLEAN, default: FALSE (i.e., CPU).
    Flag to enable GPU support for training neural network.
    The number of GPUs to use is determined by the parameters
    passed to the preprocessor.

    @note
    We have seen some memory related issues when segments
    share GPU resources.
    For example, if you provide 1 GPU and your
    database cluster is set up to have 4
    segments per segment host, it means that all 4
    segments on a segment host will share the same
    GPU. The current recommended
    configuration is 1 GPU per segment.
  </DD>

  <DT>class_values (optional)</DT>
  <DD>TEXT[], default: NULL.
    List of class labels that were used while training the model. See the 'output_table'
    column above for more details.

    @note
    If you specify the class values parameter,
    it must reflect how the dependent variable was 1-hot encoded for training. If you accidently
    pick another order that does not match the 1-hot encoding, the predictions would be wrong.
  </DD>

  <DT>normalizing_const (optional)</DT>
  <DD>DOUBLE PRECISION, default: 1.0.
  The normalizing constant to divide each value in the 'independent_varname'
  array by. For example, you would use 255 for this value if the image data is
  in the form 0-255.
  </DD>
</DL>


@anchor example
@par Examples

@note
Deep learning works best on very large datasets,
but that is not convenient for a quick introduction
to the syntax. So in this example we use an MLP on the well
known iris data set from https://archive.ics.uci.edu/ml/datasets/iris.
For more realistic examples with images please refer
to the deep learning notebooks
at https://github.com/apache/madlib-site/tree/asf-site/community-artifacts.

<h4>Classification</h4>

-#  Create an input data set.
<pre class="example">
DROP TABLE IF EXISTS iris_data;
CREATE TABLE iris_data(
    id serial,
    attributes numeric[],
    class_text varchar
);
INSERT INTO iris_data(id, attributes, class_text) VALUES
(1,ARRAY[5.1,3.5,1.4,0.2],'Iris-setosa'),
(2,ARRAY[4.9,3.0,1.4,0.2],'Iris-setosa'),
(3,ARRAY[4.7,3.2,1.3,0.2],'Iris-setosa'),
(4,ARRAY[4.6,3.1,1.5,0.2],'Iris-setosa'),
(5,ARRAY[5.0,3.6,1.4,0.2],'Iris-setosa'),
(6,ARRAY[5.4,3.9,1.7,0.4],'Iris-setosa'),
(7,ARRAY[4.6,3.4,1.4,0.3],'Iris-setosa'),
(8,ARRAY[5.0,3.4,1.5,0.2],'Iris-setosa'),
(9,ARRAY[4.4,2.9,1.4,0.2],'Iris-setosa'),
(10,ARRAY[4.9,3.1,1.5,0.1],'Iris-setosa'),
(11,ARRAY[5.4,3.7,1.5,0.2],'Iris-setosa'),
(12,ARRAY[4.8,3.4,1.6,0.2],'Iris-setosa'),
(13,ARRAY[4.8,3.0,1.4,0.1],'Iris-setosa'),
(14,ARRAY[4.3,3.0,1.1,0.1],'Iris-setosa'),
(15,ARRAY[5.8,4.0,1.2,0.2],'Iris-setosa'),
(16,ARRAY[5.7,4.4,1.5,0.4],'Iris-setosa'),
(17,ARRAY[5.4,3.9,1.3,0.4],'Iris-setosa'),
(18,ARRAY[5.1,3.5,1.4,0.3],'Iris-setosa'),
(19,ARRAY[5.7,3.8,1.7,0.3],'Iris-setosa'),
(20,ARRAY[5.1,3.8,1.5,0.3],'Iris-setosa'),
(21,ARRAY[5.4,3.4,1.7,0.2],'Iris-setosa'),
(22,ARRAY[5.1,3.7,1.5,0.4],'Iris-setosa'),
(23,ARRAY[4.6,3.6,1.0,0.2],'Iris-setosa'),
(24,ARRAY[5.1,3.3,1.7,0.5],'Iris-setosa'),
(25,ARRAY[4.8,3.4,1.9,0.2],'Iris-setosa'),
(26,ARRAY[5.0,3.0,1.6,0.2],'Iris-setosa'),
(27,ARRAY[5.0,3.4,1.6,0.4],'Iris-setosa'),
(28,ARRAY[5.2,3.5,1.5,0.2],'Iris-setosa'),
(29,ARRAY[5.2,3.4,1.4,0.2],'Iris-setosa'),
(30,ARRAY[4.7,3.2,1.6,0.2],'Iris-setosa'),
(31,ARRAY[4.8,3.1,1.6,0.2],'Iris-setosa'),
(32,ARRAY[5.4,3.4,1.5,0.4],'Iris-setosa'),
(33,ARRAY[5.2,4.1,1.5,0.1],'Iris-setosa'),
(34,ARRAY[5.5,4.2,1.4,0.2],'Iris-setosa'),
(35,ARRAY[4.9,3.1,1.5,0.1],'Iris-setosa'),
(36,ARRAY[5.0,3.2,1.2,0.2],'Iris-setosa'),
(37,ARRAY[5.5,3.5,1.3,0.2],'Iris-setosa'),
(38,ARRAY[4.9,3.1,1.5,0.1],'Iris-setosa'),
(39,ARRAY[4.4,3.0,1.3,0.2],'Iris-setosa'),
(40,ARRAY[5.1,3.4,1.5,0.2],'Iris-setosa'),
(41,ARRAY[5.0,3.5,1.3,0.3],'Iris-setosa'),
(42,ARRAY[4.5,2.3,1.3,0.3],'Iris-setosa'),
(43,ARRAY[4.4,3.2,1.3,0.2],'Iris-setosa'),
(44,ARRAY[5.0,3.5,1.6,0.6],'Iris-setosa'),
(45,ARRAY[5.1,3.8,1.9,0.4],'Iris-setosa'),
(46,ARRAY[4.8,3.0,1.4,0.3],'Iris-setosa'),
(47,ARRAY[5.1,3.8,1.6,0.2],'Iris-setosa'),
(48,ARRAY[4.6,3.2,1.4,0.2],'Iris-setosa'),
(49,ARRAY[5.3,3.7,1.5,0.2],'Iris-setosa'),
(50,ARRAY[5.0,3.3,1.4,0.2],'Iris-setosa'),
(51,ARRAY[7.0,3.2,4.7,1.4],'Iris-versicolor'),
(52,ARRAY[6.4,3.2,4.5,1.5],'Iris-versicolor'),
(53,ARRAY[6.9,3.1,4.9,1.5],'Iris-versicolor'),
(54,ARRAY[5.5,2.3,4.0,1.3],'Iris-versicolor'),
(55,ARRAY[6.5,2.8,4.6,1.5],'Iris-versicolor'),
(56,ARRAY[5.7,2.8,4.5,1.3],'Iris-versicolor'),
(57,ARRAY[6.3,3.3,4.7,1.6],'Iris-versicolor'),
(58,ARRAY[4.9,2.4,3.3,1.0],'Iris-versicolor'),
(59,ARRAY[6.6,2.9,4.6,1.3],'Iris-versicolor'),
(60,ARRAY[5.2,2.7,3.9,1.4],'Iris-versicolor'),
(61,ARRAY[5.0,2.0,3.5,1.0],'Iris-versicolor'),
(62,ARRAY[5.9,3.0,4.2,1.5],'Iris-versicolor'),
(63,ARRAY[6.0,2.2,4.0,1.0],'Iris-versicolor'),
(64,ARRAY[6.1,2.9,4.7,1.4],'Iris-versicolor'),
(65,ARRAY[5.6,2.9,3.6,1.3],'Iris-versicolor'),
(66,ARRAY[6.7,3.1,4.4,1.4],'Iris-versicolor'),
(67,ARRAY[5.6,3.0,4.5,1.5],'Iris-versicolor'),
(68,ARRAY[5.8,2.7,4.1,1.0],'Iris-versicolor'),
(69,ARRAY[6.2,2.2,4.5,1.5],'Iris-versicolor'),
(70,ARRAY[5.6,2.5,3.9,1.1],'Iris-versicolor'),
(71,ARRAY[5.9,3.2,4.8,1.8],'Iris-versicolor'),
(72,ARRAY[6.1,2.8,4.0,1.3],'Iris-versicolor'),
(73,ARRAY[6.3,2.5,4.9,1.5],'Iris-versicolor'),
(74,ARRAY[6.1,2.8,4.7,1.2],'Iris-versicolor'),
(75,ARRAY[6.4,2.9,4.3,1.3],'Iris-versicolor'),
(76,ARRAY[6.6,3.0,4.4,1.4],'Iris-versicolor'),
(77,ARRAY[6.8,2.8,4.8,1.4],'Iris-versicolor'),
(78,ARRAY[6.7,3.0,5.0,1.7],'Iris-versicolor'),
(79,ARRAY[6.0,2.9,4.5,1.5],'Iris-versicolor'),
(80,ARRAY[5.7,2.6,3.5,1.0],'Iris-versicolor'),
(81,ARRAY[5.5,2.4,3.8,1.1],'Iris-versicolor'),
(82,ARRAY[5.5,2.4,3.7,1.0],'Iris-versicolor'),
(83,ARRAY[5.8,2.7,3.9,1.2],'Iris-versicolor'),
(84,ARRAY[6.0,2.7,5.1,1.6],'Iris-versicolor'),
(85,ARRAY[5.4,3.0,4.5,1.5],'Iris-versicolor'),
(86,ARRAY[6.0,3.4,4.5,1.6],'Iris-versicolor'),
(87,ARRAY[6.7,3.1,4.7,1.5],'Iris-versicolor'),
(88,ARRAY[6.3,2.3,4.4,1.3],'Iris-versicolor'),
(89,ARRAY[5.6,3.0,4.1,1.3],'Iris-versicolor'),
(90,ARRAY[5.5,2.5,4.0,1.3],'Iris-versicolor'),
(91,ARRAY[5.5,2.6,4.4,1.2],'Iris-versicolor'),
(92,ARRAY[6.1,3.0,4.6,1.4],'Iris-versicolor'),
(93,ARRAY[5.8,2.6,4.0,1.2],'Iris-versicolor'),
(94,ARRAY[5.0,2.3,3.3,1.0],'Iris-versicolor'),
(95,ARRAY[5.6,2.7,4.2,1.3],'Iris-versicolor'),
(96,ARRAY[5.7,3.0,4.2,1.2],'Iris-versicolor'),
(97,ARRAY[5.7,2.9,4.2,1.3],'Iris-versicolor'),
(98,ARRAY[6.2,2.9,4.3,1.3],'Iris-versicolor'),
(99,ARRAY[5.1,2.5,3.0,1.1],'Iris-versicolor'),
(100,ARRAY[5.7,2.8,4.1,1.3],'Iris-versicolor'),
(101,ARRAY[6.3,3.3,6.0,2.5],'Iris-virginica'),
(102,ARRAY[5.8,2.7,5.1,1.9],'Iris-virginica'),
(103,ARRAY[7.1,3.0,5.9,2.1],'Iris-virginica'),
(104,ARRAY[6.3,2.9,5.6,1.8],'Iris-virginica'),
(105,ARRAY[6.5,3.0,5.8,2.2],'Iris-virginica'),
(106,ARRAY[7.6,3.0,6.6,2.1],'Iris-virginica'),
(107,ARRAY[4.9,2.5,4.5,1.7],'Iris-virginica'),
(108,ARRAY[7.3,2.9,6.3,1.8],'Iris-virginica'),
(109,ARRAY[6.7,2.5,5.8,1.8],'Iris-virginica'),
(110,ARRAY[7.2,3.6,6.1,2.5],'Iris-virginica'),
(111,ARRAY[6.5,3.2,5.1,2.0],'Iris-virginica'),
(112,ARRAY[6.4,2.7,5.3,1.9],'Iris-virginica'),
(113,ARRAY[6.8,3.0,5.5,2.1],'Iris-virginica'),
(114,ARRAY[5.7,2.5,5.0,2.0],'Iris-virginica'),
(115,ARRAY[5.8,2.8,5.1,2.4],'Iris-virginica'),
(116,ARRAY[6.4,3.2,5.3,2.3],'Iris-virginica'),
(117,ARRAY[6.5,3.0,5.5,1.8],'Iris-virginica'),
(118,ARRAY[7.7,3.8,6.7,2.2],'Iris-virginica'),
(119,ARRAY[7.7,2.6,6.9,2.3],'Iris-virginica'),
(120,ARRAY[6.0,2.2,5.0,1.5],'Iris-virginica'),
(121,ARRAY[6.9,3.2,5.7,2.3],'Iris-virginica'),
(122,ARRAY[5.6,2.8,4.9,2.0],'Iris-virginica'),
(123,ARRAY[7.7,2.8,6.7,2.0],'Iris-virginica'),
(124,ARRAY[6.3,2.7,4.9,1.8],'Iris-virginica'),
(125,ARRAY[6.7,3.3,5.7,2.1],'Iris-virginica'),
(126,ARRAY[7.2,3.2,6.0,1.8],'Iris-virginica'),
(127,ARRAY[6.2,2.8,4.8,1.8],'Iris-virginica'),
(128,ARRAY[6.1,3.0,4.9,1.8],'Iris-virginica'),
(129,ARRAY[6.4,2.8,5.6,2.1],'Iris-virginica'),
(130,ARRAY[7.2,3.0,5.8,1.6],'Iris-virginica'),
(131,ARRAY[7.4,2.8,6.1,1.9],'Iris-virginica'),
(132,ARRAY[7.9,3.8,6.4,2.0],'Iris-virginica'),
(133,ARRAY[6.4,2.8,5.6,2.2],'Iris-virginica'),
(134,ARRAY[6.3,2.8,5.1,1.5],'Iris-virginica'),
(135,ARRAY[6.1,2.6,5.6,1.4],'Iris-virginica'),
(136,ARRAY[7.7,3.0,6.1,2.3],'Iris-virginica'),
(137,ARRAY[6.3,3.4,5.6,2.4],'Iris-virginica'),
(138,ARRAY[6.4,3.1,5.5,1.8],'Iris-virginica'),
(139,ARRAY[6.0,3.0,4.8,1.8],'Iris-virginica'),
(140,ARRAY[6.9,3.1,5.4,2.1],'Iris-virginica'),
(141,ARRAY[6.7,3.1,5.6,2.4],'Iris-virginica'),
(142,ARRAY[6.9,3.1,5.1,2.3],'Iris-virginica'),
(143,ARRAY[5.8,2.7,5.1,1.9],'Iris-virginica'),
(144,ARRAY[6.8,3.2,5.9,2.3],'Iris-virginica'),
(145,ARRAY[6.7,3.3,5.7,2.5],'Iris-virginica'),
(146,ARRAY[6.7,3.0,5.2,2.3],'Iris-virginica'),
(147,ARRAY[6.3,2.5,5.0,1.9],'Iris-virginica'),
(148,ARRAY[6.5,3.0,5.2,2.0],'Iris-virginica'),
(149,ARRAY[6.2,3.4,5.4,2.3],'Iris-virginica'),
(150,ARRAY[5.9,3.0,5.1,1.8],'Iris-virginica');
</pre>
Create a test/validation dataset from the training data:
<pre class="example">
DROP TABLE IF EXISTS iris_train, iris_test;
-- Set seed so results are reproducible
SELECT setseed(0);
SELECT madlib.train_test_split('iris_data',     -- Source table
                               'iris',          -- Output table root name
                                0.8,            -- Train proportion
                                NULL,           -- Test proportion (0.2)
                                NULL,           -- Strata definition
                                NULL,           -- Output all columns
                                NULL,           -- Sample without replacement
                                TRUE            -- Separate output tables
                              );
SELECT COUNT(*) FROM iris_train;
</pre>
<pre class="result">
 count
------+
   120
</pre>

-# Call the preprocessor for deep learning.  For the training dataset:
<pre class="example">
\\x off
DROP TABLE IF EXISTS iris_train_packed, iris_train_packed_summary;
SELECT madlib.training_preprocessor_dl('iris_train',         -- Source table
                                       'iris_train_packed',  -- Output table
                                       'class_text',         -- Dependent variable
                                       'attributes'          -- Independent variable
                                        );
SELECT * FROM iris_train_packed_summary;
</pre>
<pre class="result">
-[ RECORD 1 ]-------+---------------------------------------------
source_table        | iris_train
output_table        | iris_train_packed
dependent_varname   | class_text
independent_varname | attributes
dependent_vartype   | character varying
class_values        | {Iris-setosa,Iris-versicolor,Iris-virginica}
buffer_size         | 60
normalizing_const   | 1.0
num_classes         | 3
</pre>
For the validation dataset:
<pre class="example">
DROP TABLE IF EXISTS iris_test_packed, iris_test_packed_summary;
SELECT madlib.validation_preprocessor_dl('iris_test',          -- Source table
                                         'iris_test_packed',   -- Output table
                                         'class_text',         -- Dependent variable
                                         'attributes',         -- Independent variable
                                         'iris_train_packed'   -- From training preprocessor step
                                          );
SELECT * FROM iris_test_packed_summary;
</pre>
<pre class="result">
-[ RECORD 1 ]-------+---------------------------------------------
source_table        | iris_test
output_table        | iris_test_packed
dependent_varname   | class_text
independent_varname | attributes
dependent_vartype   | character varying
class_values        | {Iris-setosa,Iris-versicolor,Iris-virginica}
buffer_size         | 15
normalizing_const   | 1.0
num_classes         | 3
</pre>

-# Define and load model architecture.  Use Keras to define
the model architecture:
<pre class="example">
import keras
from keras.models import Sequential
from keras.layers import Dense
model_simple = Sequential()
model_simple.add(Dense(10, activation='relu', input_shape=(4,)))
model_simple.add(Dense(10, activation='relu'))
model_simple.add(Dense(3, activation='softmax'))
model_simple.summary()
\verbatim

_________________________________________________________________
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
\endverbatim
</pre>
Export the model to JSON:
<pre class="example">
model_simple.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>
Load into model 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
                               'A simple model'       -- Descr
);
</pre>

-# Train model and view summary table:
<pre class="example">
DROP TABLE IF EXISTS iris_model, iris_model_summary;
SELECT madlib.madlib_keras_fit('iris_train_packed',   -- source table
                               'iris_model',          -- model output table
                               'model_arch_library',  -- model arch table
                                1,                    -- model arch id
                                $$ loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'] $$,  -- compile_params
                                $$ batch_size=5, epochs=3 $$,  -- fit_params
                                10                    -- num_iterations
                              );
SELECT * FROM iris_model_summary;
</pre>
<pre class="result">
-[ RECORD 1 ]-------------+--------------------------------------------------------------------------
source_table              | iris_train_packed
model                     | iris_model
dependent_varname         | class_text
independent_varname       | attributes
model_arch_table          | model_arch_library
model_id                  | 1
compile_params            |  loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']
fit_params                |  batch_size=5, epochs=3
num_iterations            | 10
validation_table          |
metrics_compute_frequency | 10
name                      |
description               |
model_type                | madlib_keras
model_size                | 0.7900390625
start_training_time       | 2019-06-05 20:55:15.785034
end_training_time         | 2019-06-05 20:55:25.373035
metrics_elapsed_time      | {9.58799290657043}
madlib_version            | 1.16
num_classes               | 3
class_values              | {Iris-setosa,Iris-versicolor,Iris-virginica}
dependent_vartype         | character varying
normalizing_const         | 1
metrics_type              | {accuracy}
training_metrics_final    | 0.766666650772
training_loss_final       | 0.721103310585
training_metrics          | {0.766666650772095}
training_loss             | {0.721103310585022}
validation_metrics_final  |
validation_loss_final     |
validation_metrics        |
validation_loss           |
metrics_iters             | {10}
</pre>

-# Use the test dataset to evaluate the model we built above:
<pre class="example">
DROP TABLE IF EXISTS iris_validate;
SELECT madlib.madlib_keras_evaluate('iris_model',       -- model
                                   'iris_test_packed',  -- test table
                                   'iris_validate'      -- output table
                                   );
SELECT * FROM iris_validate;
</pre>
<pre class="result">
       loss        |      metric       | metrics_type
-------------------+-------------------+--------------
 0.719491899013519 | 0.800000011920929 | {accuracy}
(1 row)
</pre>

-# Predict.  We will use the validation dataset for prediction
as well, which is not usual but serves to show the
syntax. The prediction is in the 'estimated_class_text'
column:
<pre class="example">
DROP TABLE IF EXISTS iris_predict;
SELECT madlib.madlib_keras_predict('iris_model', -- model
                                   'iris_test',  -- test_table
                                   'id',  -- id column
                                   'attributes', -- independent var
                                   'iris_predict'  -- output table
                                   );
SELECT * FROM iris_predict ORDER BY id;
</pre>
<pre class="result">
 id  | estimated_class_text
-----+----------------------
   4 | Iris-setosa
   6 | Iris-setosa
   8 | Iris-setosa
  12 | Iris-setosa
  13 | Iris-setosa
  15 | Iris-setosa
  24 | Iris-setosa
  30 | Iris-setosa
  38 | Iris-setosa
  49 | Iris-setosa
  60 | Iris-virginica
  68 | Iris-versicolor
  69 | Iris-versicolor
  76 | Iris-versicolor
  78 | Iris-versicolor
  81 | Iris-versicolor
  85 | Iris-virginica
  90 | Iris-versicolor
  91 | Iris-versicolor
  94 | Iris-virginica
 104 | Iris-virginica
 106 | Iris-versicolor
 107 | Iris-virginica
 110 | Iris-virginica
 119 | Iris-versicolor
 127 | Iris-virginica
 129 | Iris-virginica
 134 | Iris-versicolor
 139 | Iris-virginica
 144 | Iris-virginica
(30 rows)
</pre>
Count missclassifications:
<pre class="example">
SELECT COUNT(*) FROM iris_predict JOIN iris_test USING (id)
WHERE iris_predict.estimated_class_text != iris_test.class_text;
</pre>
<pre class="result">
 count
-------+
     6
(1 row)
</pre>
Accuracy:
<pre class="example">
SELECT round(count(*)*100/(150*0.2),2) as test_accuracy_percent from
    (select iris_test.class_text as actual, iris_predict.estimated_class_text as estimated
     from iris_predict inner join iris_test
     on iris_test.id=iris_predict.id) q
WHERE q.actual=q.estimated;
</pre>
<pre class="result">
 test_accuracy_percent
-----------------------+
                 80.00
(1 row)
</pre>

-# Predict BYOM.
We will use the validation dataset for prediction
as well, which is not usual but serves to show the
syntax. See <a href="group__grp__keras__model__arch.html">load_keras_model</a>
for details on how to load the model architecture and weights.
In this example we will use weights we already have:
<pre class="example">
UPDATE model_arch_library
SET model_weights = iris_model.model_weights
FROM iris_model
WHERE model_arch_library.model_id = 1;
</pre>
Now train using a model from the model architecture table directly
without referencing the model table from the MADlib training.  Note that if you
specify the class values parameter as we do below, it must reflect how the dependent
variable was 1-hot encoded for training. In this example the 'training_preprocessor_dl()'
in Step 2 above encoded in the order {'Iris-setosa', 'Iris-versicolor', 'Iris-virginica'} so
this is the order we pass in the parameter. If we accidently pick another order that does
not match the 1-hot encoding, the predictions would be wrong.
<pre class="example">
DROP TABLE IF EXISTS iris_predict_byom;
SELECT madlib.madlib_keras_predict_byom('model_arch_library',  -- model arch table
                                         1,                    -- model arch id
                                        'iris_test',           -- test_table
                                        'id',                  -- id column
                                        'attributes',          -- independent var
                                        'iris_predict_byom',   -- output table
                                        'response',            -- prediction type
                                         FALSE,                -- use GPUs
                                         ARRAY['Iris-setosa', 'Iris-versicolor', 'Iris-virginica'], -- class values
                                         1.0                   -- normalizing const
                                   );
SELECT * FROM iris_predict_byom ORDER BY id;
</pre>
The prediction is in the 'estimated_dependent_var' column:
<pre class="result">
 id  | estimated_dependent_var
-----+----------------------
   4 | Iris-setosa
   6 | Iris-setosa
   8 | Iris-setosa
  12 | Iris-setosa
  13 | Iris-setosa
  15 | Iris-setosa
  24 | Iris-setosa
  30 | Iris-setosa
  38 | Iris-setosa
  49 | Iris-setosa
  60 | Iris-virginica
  68 | Iris-versicolor
  69 | Iris-versicolor
  76 | Iris-versicolor
  78 | Iris-versicolor
  81 | Iris-versicolor
  85 | Iris-virginica
  90 | Iris-versicolor
  91 | Iris-versicolor
  94 | Iris-virginica
 104 | Iris-virginica
 106 | Iris-versicolor
 107 | Iris-virginica
 110 | Iris-virginica
 119 | Iris-versicolor
 127 | Iris-virginica
 129 | Iris-virginica
 134 | Iris-versicolor
 139 | Iris-virginica
 144 | Iris-virginica
(30 rows)
 </pre>
Count missclassifications:
<pre class="example">
SELECT COUNT(*) FROM iris_predict_byom JOIN iris_test USING (id)
WHERE iris_predict_byom.estimated_dependent_var != iris_test.class_text;
</pre>
<pre class="result">
 count
-------+
     6
(1 row)
</pre>
Accuracy:
<pre class="example">
SELECT round(count(*)*100/(150*0.2),2) as test_accuracy_percent from
    (select iris_test.class_text as actual, iris_predict_byom.estimated_dependent_var as estimated
     from iris_predict_byom inner join iris_test
     on iris_test.id=iris_predict_byom.id) q
WHERE q.actual=q.estimated;
</pre>
<pre class="result">
 test_accuracy_percent
-----------------------+
                 80.00
(1 row)
</pre>


<h4>Classification with Other Parameters</h4>

-# Validation dataset.  Now use a validation dataset
and compute metrics every 3rd iteration using
the 'metrics_compute_frequency' parameter. This can
help reduce run time if you do not need metrics
computed at every iteration.
<pre class="example">
DROP TABLE IF EXISTS iris_model, iris_model_summary;
SELECT madlib.madlib_keras_fit('iris_train_packed',   -- source table
                               'iris_model',          -- model output table
                               'model_arch_library',  -- model arch table
                                1,                    -- model arch id
                                $$ loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'] $$,  -- compile_params
                                $$ batch_size=5, epochs=3 $$,  -- fit_params
                                10,                   -- num_iterations
                                FALSE,                -- use GPUs
                                'iris_test_packed',   -- validation dataset
                                3,                    -- metrics compute frequency
                                FALSE,                -- warm start
                               'Sophie L.',           -- name
                               'Simple MLP for iris dataset'  -- description
                              );
SELECT * FROM iris_model_summary;
</pre>
<pre class="result">
-[ RECORD 1 ]-------------+--------------------------------------------------------------------------
source_table              | iris_train_packed
model                     | iris_model
dependent_varname         | class_text
independent_varname       | attributes
model_arch_table          | model_arch_library
model_id                  | 1
compile_params            |  loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']
fit_params                |  batch_size=5, epochs=3
num_iterations            | 10
validation_table          | iris_test_packed
metrics_compute_frequency | 3
name                      | Sophie L.
description               | Simple MLP for iris dataset
model_type                | madlib_keras
model_size                | 0.7900390625
start_training_time       | 2019-06-05 20:58:23.224629
end_training_time         | 2019-06-05 20:58:35.477499
metrics_elapsed_time      | {4.69859290122986,8.2062520980835,10.8104848861694,12.2528700828552}
madlib_version            | 1.16
num_classes               | 3
class_values              | {Iris-setosa,Iris-versicolor,Iris-virginica}
dependent_vartype         | character varying
normalizing_const         | 1
metrics_type              | {accuracy}
training_metrics_final    | 0.941666662693
training_loss_final       | 0.40586027503
training_metrics          | {0.699999988079071,0.800000011920929,0.899999976158142,0.941666662693024}
training_loss             | {0.825238645076752,0.534248650074005,0.427499741315842,0.405860275030136}
validation_metrics_final  | 0.866666674614
validation_loss_final     | 0.409001916647
validation_metrics        | {0.733333349227905,0.733333349227905,0.866666674613953,0.866666674613953}
validation_loss           | {0.827081918716431,0.536275088787079,0.431326270103455,0.409001916646957}
metrics_iters             | {3,6,9,10}
</pre>

-# Predict probabilities for each class:
<pre class="example">
DROP TABLE IF EXISTS iris_predict;
SELECT madlib.madlib_keras_predict('iris_model',      -- model
                                   'iris_test',       -- test_table
                                   'id',              -- id column
                                   'attributes',      -- independent var
                                   'iris_predict',    -- output table
                                   'prob'             -- response type
                                   );
SELECT * FROM iris_predict ORDER BY id;
</pre>
<pre class="result">
 id  | prob_Iris-setosa | prob_Iris-versicolor | prob_Iris-virginica
-----+------------------+----------------------+---------------------
   4 |        0.9241953 |          0.059390426 |          0.01641435
   6 |        0.9657151 |           0.02809224 |        0.0061926916
   8 |        0.9543316 |           0.03670931 |         0.008959154
  12 |       0.93851465 |          0.048681837 |         0.012803554
  13 |       0.93832576 |           0.04893658 |         0.012737647
  15 |       0.98717564 |           0.01091238 |        0.0019119986
  24 |        0.9240628 |          0.060805064 |         0.015132156
  30 |       0.92063266 |          0.062279057 |         0.017088294
  38 |        0.9353765 |          0.051353406 |         0.013270103
  49 |        0.9709265 |          0.023811856 |         0.005261566
  60 |      0.034395564 |            0.5260507 |          0.43955377
  68 |      0.031360663 |           0.53689945 |          0.43173987
  69 |     0.0098787155 |           0.46121457 |          0.52890676
  76 |      0.031186827 |            0.5644549 |          0.40435827
  78 |       0.00982633 |           0.48929632 |           0.5008774
  81 |       0.03658528 |           0.53248984 |           0.4309249
  85 |      0.015423619 |           0.48452598 |           0.5000504
  90 |      0.026857043 |            0.5155698 |          0.45757324
  91 |      0.013675574 |           0.47155368 |           0.5147708
  94 |      0.073440716 |            0.5418821 |           0.3846772
 104 |     0.0021637122 |            0.3680499 |          0.62978643
 106 |    0.00052832486 |           0.30891812 |           0.6905536
 107 |      0.007315576 |           0.40949163 |           0.5831927
 110 |     0.0022259138 |            0.4058138 |          0.59196025
 119 |    0.00018505375 |           0.24510723 |           0.7547077
 127 |      0.009542585 |           0.46958733 |          0.52087003
 129 |     0.0019719477 |           0.36288205 |            0.635146
 134 |     0.0056418083 |           0.43401477 |          0.56034344
 139 |       0.01067015 |            0.4755573 |          0.51377255
 144 |     0.0018909549 |           0.37689638 |           0.6212126
(30 rows)
</pre>

-# Warm start.  Next, use the warm_start parameter
to continue learning, using the coefficients from
the run above. Note that we don't drop the
model table or model summary table:
<pre class="example">
SELECT madlib.madlib_keras_fit('iris_train_packed',   -- source table
                               'iris_model',          -- model output table
                               'model_arch_library',  -- model arch table
                                1,                    -- model arch id
                                $$ loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'] $$,  -- compile_params
                                $$ batch_size=5, epochs=3 $$,  -- fit_params
                                5,                   -- num_iterations
                                FALSE,               -- use GPUs
                                'iris_test_packed',   -- validation dataset
                                1,                    -- metrics compute frequency
                                TRUE,                 -- warm start
                               'Sophie L.',           -- name
                               'Simple MLP for iris dataset'  -- description
                              );
SELECT * FROM iris_model_summary;
</pre>
<pre class="result">
-[ RECORD 1 ]-------------+--------------------------------------------------------------------------------------------
source_table              | iris_train_packed
model                     | iris_model
dependent_varname         | class_text
independent_varname       | attributes
model_arch_table          | model_arch_library
model_id                  | 1
compile_params            |  loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']
fit_params                |  batch_size=5, epochs=3
num_iterations            | 5
validation_table          | iris_test_packed
metrics_compute_frequency | 1
name                      | Sophie L.
description               | Simple MLP for iris dataset
model_type                | madlib_keras
model_size                | 0.7900390625
start_training_time       | 2019-06-05 20:59:43.971792
end_training_time         | 2019-06-05 20:59:51.654586
metrics_elapsed_time      | {2.89326310157776,4.14273309707642,5.24781513214111,6.34498596191406,7.68279695510864}
madlib_version            | 1.16
num_classes               | 3
class_values              | {Iris-setosa,Iris-versicolor,Iris-virginica}
dependent_vartype         | character varying
normalizing_const         | 1
metrics_type              | {accuracy}
training_metrics_final    | 0.933333337307
training_loss_final       | 0.334455043077
training_metrics          | {0.933333337306976,0.933333337306976,0.975000023841858,0.975000023841858,0.933333337306976}
training_loss             | {0.386842548847198,0.370587915182114,0.357161343097687,0.344598710536957,0.334455043077469}
validation_metrics_final  | 0.866666674614
validation_loss_final     | 0.34414178133
validation_metrics        | {0.866666674613953,0.866666674613953,0.933333337306976,0.866666674613953,0.866666674613953}
validation_loss           | {0.391442179679871,0.376414686441422,0.362262904644012,0.351912915706635,0.344141781330109}
metrics_iters             | {1,2,3,4,5}
</pre>
Note that the loss and accuracy values pick up from where the previous run left off.

<h4>Transfer Learning</h4>
Here we want
to start with initial weights from a pre-trained model
rather than training from scratch. We also want to use
a model architecture with the earlier feature layer(s)
frozen to save on training time. The example below is
somewhat contrived but gives you the idea of the steps.

-# Define and load a model architecture with
the 1st hidden layer frozen:

<pre class="example">
model_transfer = Sequential()
model_transfer.add(Dense(10, activation='relu', input_shape=(4,), trainable=False))
model_transfer.add(Dense(10, activation='relu'))
model_transfer.add(Dense(3, activation='softmax'))
model_simple.summary()
\verbatim

_________________________________________________________________
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: 143
Non-trainable params: 50
\endverbatim
</pre>
Export the model to JSON:
<pre class="example">
model_simple.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_2", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "dtype": "float32", "activation": "relu", "trainable": false, "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_3", "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_4", "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>
Load into model architecture table:
<pre class="example">
SELECT madlib.load_keras_model('model_arch_library',  -- Output table,
$$
{"class_name": "Sequential", "keras_version": "2.1.6", "config": [{"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_2", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "dtype": "float32", "activation": "relu", "trainable": false, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "batch_input_shape": [null, 4], "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_3", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "use_bias": true, "activity_regularizer": null}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "VarianceScaling", "config": {"distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg"}}, "name": "dense_4", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "activation": "softmax", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 3, "use_bias": true, "activity_regularizer": null}}], "backend": "tensorflow"}
$$
::json,  -- JSON blob
                               NULL,                  -- Weights
                               'Maria',               -- Name
                               'A transfer model'     -- Descr
);
</pre>
Fetch the weights from a previous MADlib run. (Normally
these would be downloaded from a source that trained
the same model architecture on a related dataset.)
<pre class="example">
UPDATE model_arch_library
SET model_weights = iris_model.model_weights
FROM iris_model
WHERE model_arch_library.model_id = 2;
</pre>
Now train the model using the transfer model and the pre-trained weights:
<pre class="example">
DROP TABLE IF EXISTS iris_model, iris_model_summary;
SELECT madlib.madlib_keras_fit('iris_train_packed',   -- source table
                               'iris_model',          -- model output table
                               'model_arch_library',  -- model arch table
                                2,                    -- model arch id
                                $$ loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'] $$,  -- compile_params
                                $$ batch_size=5, epochs=3 $$,  -- fit_params
                                10                    -- num_iterations
                              );
SELECT * FROM iris_model_summary;
</pre>
<pre class="result">
-[ RECORD 1 ]-------------+--------------------------------------------------------------------------
source_table              | iris_train_packed
model                     | iris_model
dependent_varname         | class_text
independent_varname       | attributes
model_arch_table          | model_arch_library
model_id                  | 2
compile_params            |  loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']
fit_params                |  batch_size=5, epochs=3
num_iterations            | 10
validation_table          |
metrics_compute_frequency | 10
name                      |
description               |
model_type                | madlib_keras
model_size                | 0.7900390625
start_training_time       | 2019-06-05 21:01:03.998422
end_training_time         | 2019-06-05 21:01:13.525838
metrics_elapsed_time      | {9.52741599082947}
madlib_version            | 1.16
num_classes               | 3
class_values              | {Iris-setosa,Iris-versicolor,Iris-virginica}
dependent_vartype         | character varying
normalizing_const         | 1
metrics_type              | {accuracy}
training_metrics_final    | 0.975000023842
training_loss_final       | 0.245171800256
training_metrics          | {0.975000023841858}
training_loss             | {0.245171800255775}
validation_metrics_final  |
validation_loss_final     |
validation_metrics        |
validation_loss           |
metrics_iters             | {10}
</pre>

@anchor notes
@par Notes

1. Refer to the deep learning section of the Apache MADlib
wiki [5] for important information including supported libraries
and versions.

2. Classification is currently supported, not regression.

3. Reminder about the distinction between warm start and transfer learning.  Warm start uses model
state (weights) from the model output table from a previous training run -
set the 'warm_start' parameter to TRUE in the fit function.
Transfer learning uses initial model state (weights) stored in the 'model_arch_table' - in this case set the
'warm_start' parameter to FALSE in the fit function.

@anchor background
@par Technical Background

For an introduction to deep learning foundations, including MLP and CNN,
refer to [6].

This module trains a single large model across the database cluster
using the bulk synchronous parallel (BSP) approach, with model averaging [7].

On the effect of database cluster size: as the database cluster size increases, the per iteration
loss will be higher since the model only sees 1/n of the data, where n is the number of segments.
However, each iteration runs faster than single node because it is only traversing 1/n of the data.
For highly non-convex solution spaces, convergence behavior may diminish as cluster size increases.
Ensure that each segment has sufficient volume of data and examples of each class value.

Alternatively, to train multiple models at the same time for model
architecture search or hyperparameter tuning, you can
use <a href="group__grp__keras__run__model__selection.html">Model Selection</a>,
which does not do model averaging and hence may have better covergence efficiency.

@anchor literature
@literature

@anchor mlp-lit-1
[1] https://keras.io/

[2] https://www.tensorflow.org/

[3] "Neural Networks for Machine Learning", Lectures 6a and 6b on mini-batch gradient descent,
Geoffrey Hinton with Nitish Srivastava and Kevin Swersky,
http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf

[4] https://keras.io/models/model/

[5] Deep learning section of Apache MADlib wiki, https://cwiki.apache.org/confluence/display/MADLIB/Deep+Learning

[6] Deep Learning, Ian Goodfellow, Yoshua Bengio and Aaron Courville, MIT Press, 2016.

[7] "Resource-Efficient and Reproducible Model Selection on Deep Learning Systems," Supun Nakandala,
Yuhao Zhang, and Arun Kumar, Technical Report, Computer Science and Engineering, University of California,
San Diego https://adalabucsd.github.io/papers/TR_2019_Cerebro.pdf.

[8] Greenplum Database server configuration parameters https://gpdb.docs.pivotal.io/latest/ref_guide/config_params/guc-list.html

@anchor related
@par Related Topics

File madlib_keras.sql_in documenting the training, evaluate and predict functions.

*/

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_fit(
    source_table            VARCHAR,
    model                   VARCHAR,
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    compile_params          VARCHAR,
    fit_params              VARCHAR,
    num_iterations          INTEGER,
    use_gpus                BOOLEAN,
    validation_table        VARCHAR,
    metrics_compute_frequency  INTEGER,
    warm_start              BOOLEAN,
    name                    VARCHAR,
    description             VARCHAR
) RETURNS VOID AS $$
    PythonFunctionBodyOnly(`deep_learning', `madlib_keras')
    from utilities.control import SetGUC
    with AOControl(False):
        with SetGUC("plan_cache_mode", "force_generic_plan"):
            madlib_keras.fit(**globals())
$$ LANGUAGE plpythonu VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_fit(
    source_table            VARCHAR,
    model                   VARCHAR,
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    compile_params          VARCHAR,
    fit_params              VARCHAR,
    num_iterations          INTEGER,
    use_gpus                BOOLEAN,
    validation_table        VARCHAR,
    metrics_compute_frequency  INTEGER,
    warm_start              BOOLEAN,
    name                    VARCHAR
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_fit($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_fit(
    source_table            VARCHAR,
    model                   VARCHAR,
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    compile_params          VARCHAR,
    fit_params              VARCHAR,
    num_iterations          INTEGER,
    use_gpus                BOOLEAN,
    validation_table        VARCHAR,
    metrics_compute_frequency  INTEGER,
    warm_start              BOOLEAN
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_fit($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, NULL, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');


CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_fit(
    source_table            VARCHAR,
    model                   VARCHAR,
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    compile_params          VARCHAR,
    fit_params              VARCHAR,
    num_iterations          INTEGER,
    use_gpus                BOOLEAN,
    validation_table        VARCHAR,
    metrics_compute_frequency  INTEGER
) RETURNS VOID AS $$
SELECT MADLIB_SCHEMA.madlib_keras_fit($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, NULL, NULL, NULL);
$$ LANGUAGE sql VOLATILE
    m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');


CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_fit(
    source_table            VARCHAR,
    model                   VARCHAR,
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    compile_params          VARCHAR,
    fit_params              VARCHAR,
    num_iterations          INTEGER,
    use_gpus                BOOLEAN,
    validation_table        VARCHAR
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_fit($1, $2, $3, $4, $5, $6, $7, $8, $9, NULL, NULL, NULL, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_fit(
    source_table            VARCHAR,
    model                   VARCHAR,
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    compile_params          VARCHAR,
    fit_params              VARCHAR,
    num_iterations          INTEGER,
    use_gpus                BOOLEAN
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_fit($1, $2, $3, $4, $5, $6, $7, $8, NULL, NULL, NULL, NULL, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_fit(
    source_table            VARCHAR,
    model                   VARCHAR,
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    compile_params          VARCHAR,
    fit_params              VARCHAR,
    num_iterations          INTEGER
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_fit($1, $2, $3, $4, $5, $6, $7, FALSE, NULL, NULL, NULL, NULL, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.fit_transition(
    state                       BYTEA,
    dependent_var               BYTEA,
    independent_var             BYTEA,
    dependent_var_shape         INTEGER[],
    independent_var_shape       INTEGER[],
    model_architecture          TEXT,
    compile_params              TEXT,
    fit_params                  TEXT,
    dist_key                    INTEGER,
    dist_key_mapping            INTEGER[],
    current_seg_id              INTEGER,
    segments_per_host           INTEGER,
    images_per_seg              INTEGER[],
    use_gpus                    BOOLEAN,
    accessible_gpus_for_seg                INTEGER[],
    prev_serialized_weights     BYTEA,
    is_final_iteration          BOOLEAN
) RETURNS BYTEA AS $$
PythonFunctionBodyOnlyNoSchema(`deep_learning', `madlib_keras')
    return madlib_keras.fit_transition(**globals())
$$ LANGUAGE plpythonu
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `NO SQL', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.fit_merge(
    state1          BYTEA,
    state2          BYTEA
) RETURNS BYTEA AS $$
PythonFunctionBodyOnlyNoSchema(`deep_learning', `madlib_keras')
    return madlib_keras.fit_merge(**globals())
$$ LANGUAGE plpythonu
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `NO SQL', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.fit_final(
    state BYTEA
) RETURNS BYTEA AS $$
PythonFunctionBodyOnlyNoSchema(`deep_learning', `madlib_keras')
    return madlib_keras.fit_final(**globals())
$$ LANGUAGE plpythonu
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `NO SQL', `');

DROP AGGREGATE IF EXISTS MADLIB_SCHEMA.fit_step(
    BYTEA,
    BYTEA,
    TEXT,
    TEXT,
    TEXT,
    TEXT,
    TEXT,
    INTEGER,
    INTEGER[],
    INTEGER,
    INTEGER,
    INTEGER[],
    BOOLEAN,
    INTEGER[],
    BYTEA,
    BOOLEAN);
CREATE AGGREGATE MADLIB_SCHEMA.fit_step(
    /* dep_var */                BYTEA,
    /* ind_var */                BYTEA,
    /* dep_var_shape */          INTEGER[],
    /* ind_var_shape */          INTEGER[],
    /* model_architecture */     TEXT,
    /* compile_params */         TEXT,
    /* fit_params */             TEXT,
    /* dist_key */               INTEGER,
    /* dist_key_mapping */       INTEGER[],
    /* current_seg_id */         INTEGER,
    /* segments_per_host */      INTEGER,
    /* images_per_seg */         INTEGER[],
    /* use_gpus  */              BOOLEAN,
    /* segments_per_host  */     INTEGER[],
    /* serialized_weights */     BYTEA,
    /* is_final_iteration */     BOOLEAN
)(
    STYPE=BYTEA,
    SFUNC=MADLIB_SCHEMA.fit_transition,
    m4_ifdef(`__POSTGRESQL__', `', `prefunc=MADLIB_SCHEMA.fit_merge,')
    FINALFUNC=MADLIB_SCHEMA.fit_final
);

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict(
    model_table             VARCHAR,
    test_table              VARCHAR,
    id_col                  VARCHAR,
    independent_varname     VARCHAR,
    output_table            VARCHAR,
    pred_type               VARCHAR,
    use_gpus                BOOLEAN,
    mst_key                 INTEGER
) RETURNS VOID AS $$
    PythonFunctionBodyOnly(`deep_learning', `madlib_keras_predict')
    from utilities.control import SetGUC
    with AOControl(False):
        with SetGUC("plan_cache_mode", "force_generic_plan"):
            madlib_keras_predict.Predict(schema_madlib,
                   model_table,
                   test_table,
                   id_col,
                   independent_varname,
                   output_table,
                   pred_type,
                   use_gpus,
                   mst_key)
$$ LANGUAGE plpythonu VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict(
    model_table             VARCHAR,
    test_table              VARCHAR,
    id_col                  VARCHAR,
    independent_varname     VARCHAR,
    output_table            VARCHAR,
    pred_type               VARCHAR,
    use_gpus                BOOLEAN
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_predict($1, $2, $3, $4, $5, $6, $7, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict(
    model_table             VARCHAR,
    test_table              VARCHAR,
    id_col                  VARCHAR,
    independent_varname     VARCHAR,
    output_table            VARCHAR,
    pred_type               VARCHAR
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_predict($1, $2, $3, $4, $5, $6, FALSE, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict(
    model_table             VARCHAR,
    test_table              VARCHAR,
    id_col                  VARCHAR,
    independent_varname     VARCHAR,
    output_table            VARCHAR
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_predict($1, $2, $3, $4, $5, NULL, FALSE, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.internal_keras_predict(
    independent_var    REAL[],
    model_architecture TEXT,
    model_weights      BYTEA,
    is_response        BOOLEAN,
    normalizing_const  DOUBLE PRECISION,
    current_seg_id     INTEGER,
    seg_ids            INTEGER[],
    images_per_seg     INTEGER[],
    use_gpus           BOOLEAN,
    gpus_per_host      INTEGER,
    segments_per_host  INTEGER
) RETURNS DOUBLE PRECISION[] AS $$
    PythonFunctionBodyOnlyNoSchema(`deep_learning', `madlib_keras_predict')
    return madlib_keras_predict.internal_keras_predict(**globals())
$$ LANGUAGE plpythonu VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA', `');

-------------------------------------------------------------------------------
CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict_byom(
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    test_table              VARCHAR,
    id_col                  VARCHAR,
    independent_varname     VARCHAR,
    output_table            VARCHAR,
    pred_type               VARCHAR,
    use_gpus                BOOLEAN,
    class_values            TEXT[],
    normalizing_const       DOUBLE PRECISION
) RETURNS VOID AS $$
    PythonFunctionBodyOnly(`deep_learning', `madlib_keras_predict')
    from utilities.control import SetGUC
    with AOControl(False):
        with SetGUC("plan_cache_mode", "force_generic_plan"):
            madlib_keras_predict.PredictBYOM(**globals())
$$ LANGUAGE plpythonu VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict_byom(
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    test_table              VARCHAR,
    id_col                  VARCHAR,
    independent_varname     VARCHAR,
    output_table            VARCHAR,
    pred_type               VARCHAR,
    use_gpus                BOOLEAN,
    class_values            TEXT[]
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_predict_byom($1, $2, $3, $4, $5, $6, $7, $8, $9, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA', `');


CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict_byom(
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    test_table              VARCHAR,
    id_col                  VARCHAR,
    independent_varname     VARCHAR,
    output_table            VARCHAR,
    pred_type               VARCHAR,
    use_gpus                BOOLEAN
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_predict_byom($1, $2, $3, $4, $5, $6, $7, $8, NULL, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA', `');


CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict_byom(
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    test_table              VARCHAR,
    id_col                  VARCHAR,
    independent_varname     VARCHAR,
    output_table            VARCHAR,
    pred_type               VARCHAR
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_predict_byom($1, $2, $3, $4, $5, $6, $7, NULL, NULL, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict_byom(
    model_arch_table        VARCHAR,
    model_id                INTEGER,
    test_table              VARCHAR,
    id_col                  VARCHAR,
    independent_varname     VARCHAR,
    output_table            VARCHAR
) RETURNS VOID AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_predict_byom($1, $2, $3, $4, $5, $6, NULL, NULL, NULL, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA', `');

-------------------------------------------------------------------------------
CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_evaluate(
    model_table             VARCHAR,
    test_table              VARCHAR,
    output_table            VARCHAR,
    use_gpus                BOOLEAN,
    mst_key                 INTEGER
) RETURNS VOID AS $$
    PythonFunctionBodyOnly(`deep_learning', `madlib_keras')
    from utilities.control import SetGUC
    with AOControl(False):
        with SetGUC("plan_cache_mode", "force_generic_plan"):
            madlib_keras.evaluate(**globals())
$$ LANGUAGE plpythonu VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_evaluate(
    model_table             VARCHAR,
    test_table              VARCHAR,
    output_table            VARCHAR,
    use_gpus                BOOLEAN
) RETURNS VOID AS $$
  SELECT MADLIB_SCHEMA.madlib_keras_evaluate($1, $2, $3, $4, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_evaluate(
    model_table             VARCHAR,
    test_table              VARCHAR,
    output_table            VARCHAR
) RETURNS VOID AS $$
  SELECT MADLIB_SCHEMA.madlib_keras_evaluate($1, $2, $3, NULL, NULL);
$$ LANGUAGE sql VOLATILE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `MODIFIES SQL DATA');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.internal_keras_eval_transition(
    state                              REAL[3],
    dependent_var                      BYTEA,
    independent_var                    BYTEA,
    dependent_var_shape                INTEGER[],
    independent_var_shape              INTEGER[],
    model_architecture                 TEXT,
    serialized_weights                 BYTEA,
    compile_params                     TEXT,
    dist_key                           INTEGER,
    dist_key_mapping                   INTEGER[],
    current_seg_id                     INTEGER,
    segments_per_host                  INTEGER,
    images_per_seg                     INTEGER[],
    use_gpus                           BOOLEAN,
    accessible_gpus_for_seg                       INTEGER[],
    is_final_iteration                 BOOLEAN
) RETURNS REAL[3] AS $$
PythonFunctionBodyOnlyNoSchema(`deep_learning', `madlib_keras')
    return madlib_keras.internal_keras_eval_transition(**globals())
$$ LANGUAGE plpythonu
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `NO SQL', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.internal_keras_eval_merge(
    state1          REAL[3],
    state2          REAL[3]
) RETURNS REAL[3] AS $$
PythonFunctionBodyOnlyNoSchema(`deep_learning', `madlib_keras')
    return madlib_keras.internal_keras_eval_merge(**globals())
$$ LANGUAGE plpythonu
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `NO SQL', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.internal_keras_eval_final(
    state REAL[3]
) RETURNS REAL[2] AS $$
PythonFunctionBodyOnlyNoSchema(`deep_learning', `madlib_keras')
    return madlib_keras.internal_keras_eval_final(**globals())
$$ LANGUAGE plpythonu
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `NO SQL', `');

DROP AGGREGATE IF EXISTS MADLIB_SCHEMA.internal_keras_evaluate(
                                       BYTEA,
                                       BYTEA,
                                       INTEGER[],
                                       INTEGER[],
                                       TEXT,
                                       BYTEA,
                                       TEXT,
                                       INTEGER,
                                       INTEGER[],
                                       INTEGER,
                                       INTEGER,
                                       INTEGER[],
                                       BOOLEAN,
                                       INTEGER[],
                                       BOOLEAN);

CREATE AGGREGATE MADLIB_SCHEMA.internal_keras_evaluate(
    /* dependent_var */             BYTEA,
    /* independent_var */           BYTEA,
    /* dependent_var_shape */       INTEGER[],
    /* independent_var_shape */     INTEGER[],
    /* model_architecture */        TEXT,
    /* model_weights */             BYTEA,
    /* compile_params */            TEXT,
    /* dist_key */                  INTEGER,
    /* dist_key_mapping */          INTEGER[],
    /* current_seg_id */            INTEGER,
    /* segments_per_host */         INTEGER,
    /* images_per_seg*/             INTEGER[],
    /* use_gpus */                  BOOLEAN,
    /* accessible_gpus_for_seg */              INTEGER[],
    /* is_final_iteration */        BOOLEAN
)(
    STYPE=REAL[3],
    INITCOND='{0,0,0}',
    SFUNC=MADLIB_SCHEMA.internal_keras_eval_transition,
    m4_ifdef(`__POSTGRESQL__', `', `prefunc=MADLIB_SCHEMA.internal_keras_eval_merge,')
    FINALFUNC=MADLIB_SCHEMA.internal_keras_eval_final
);


CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_fit(
     message VARCHAR
) RETURNS VARCHAR AS $$
    PythonFunctionBodyOnly(`deep_learning', `madlib_keras')
    with AOControl(False):
        return madlib_keras.fit_help(**globals())
$$ LANGUAGE plpythonu IMMUTABLE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `CONTAINS SQL', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_fit()
RETURNS VARCHAR AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_fit('');
$$ LANGUAGE sql IMMUTABLE
m4_ifdef(`\_\_HAS_FUNCTION_PROPERTIES\_\_', `CONTAINS SQL', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_evaluate(
     message VARCHAR
) RETURNS VARCHAR AS $$
    PythonFunctionBodyOnly(`deep_learning', `madlib_keras')
    with AOControl(False):
        return madlib_keras.evaluate_help(**globals())
$$ LANGUAGE plpythonu IMMUTABLE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `CONTAINS SQL', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_evaluate()
RETURNS VARCHAR AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_evaluate('');
$$ LANGUAGE sql IMMUTABLE
m4_ifdef(`\_\_HAS_FUNCTION_PROPERTIES\_\_', `CONTAINS SQL', `');


CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict(
     message VARCHAR
) RETURNS VARCHAR AS $$
    PythonFunctionBodyOnly(`deep_learning', `madlib_keras_predict')
    with AOControl(False):
        return madlib_keras_predict.predict_help(**globals())
$$ LANGUAGE plpythonu IMMUTABLE
m4_ifdef(`__HAS_FUNCTION_PROPERTIES__', `CONTAINS SQL', `');

CREATE OR REPLACE FUNCTION MADLIB_SCHEMA.madlib_keras_predict()
RETURNS VARCHAR AS $$
    SELECT MADLIB_SCHEMA.madlib_keras_predict('');
$$ LANGUAGE sql IMMUTABLE
m4_ifdef(`\_\_HAS_FUNCTION_PROPERTIES\_\_', `CONTAINS SQL', `');