{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Use AWS Glue Databrew from within Jupyter notebooks to prepare data for ML models \n",
    "\n",
    "\n",
    "---\n",
    "\n",
    "This notebook  walks through the steps to configure and use  open source Jupyterlab extension for AWS Glue Databrew to prepare data for a sample anomaly detection model.\n",
    "\n",
    "The [electricity consumption dataset](https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014#) is used in this notebook. A subset of original dataset with 4 customer datapoints is used as a starting point. A series of DataBrew transformations are applied on the dataset to prepare it for Random Cut Forests anomaly detection model.  On the prepared dataset, a RCF model is trained and deployed in SageMaker\n",
    "\n",
    "\n",
    "Please make sure the kernel is set to 'python3'"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Install the packages needed to run this notebook"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install awswrangler\n",
    "!pip install --upgrade sagemaker\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Import the packages "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import boto3\n",
    "import sagemaker as sm\n",
    "from sagemaker import *\n",
    "import awswrangler as wr\n",
    "import matplotlib.pyplot as plt\n",
    "import os\n",
    "import pandas as pd\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### S3 bucket where the raw and transformed data will be stored and the role details"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "session = sm.Session()\n",
    "### **** 'data_bucket' should point to bucket name you are using DataBrew and model Training ***** #### \n",
    "data_bucket=session.default_bucket() \n",
    "#s3_bucket=#input_s3_bucket#\n",
    "role_arn=session.get_caller_identity_arn()\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Data Preparation using AWS Glue DataBrew"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Exploring the prepared data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "pc_processed_path=os.path.join('s3://',data_bucket,'prefix_where_DataBrew_output_is_stored')\n",
    "columns=['timestamp','client_id','hourly_consumption']\n",
    "pc_processed_df = wr.s3.read_csv(path=pc_processed_path)\n",
    "pc_processed_df=pc_processed_df [columns]\n",
    "#columns[0]='timestamp'\n",
    "#pc_processed_df.columns=columns\n",
    "pc_processed_df.client_id.unique()\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### plotting the raw timeseries electricity consumption"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "figure, axes = plt.subplots(3, 1)\n",
    "figure.set_figheight(8)\n",
    "figure.set_figwidth(15)\n",
    "pc_processed_df[pc_processed_df['client_id']=='MT_012'].plot(ax=axes[0],title='MT_012')                                                \n",
    "pc_processed_df[pc_processed_df['client_id']=='MT_013'].plot(ax=axes[1],title='MT_013')\n",
    "pc_processed_df[pc_processed_df['client_id']=='MT_132'].plot(ax=axes[2],title='MT_132')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Lets train our model with ***MT_132*** consumption data. Since RCF requires one time series and integer values, lets filter and convert the consumption data to inetger data type"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_df=pc_processed_df[(pc_processed_df['client_id']=='MT_132') & (pc_processed_df['timestamp']<'2014-11-01')]\n",
    "train_df=train_df.drop(['timestamp','client_id'],axis=1)\n",
    "train_df.hourly_consumption=train_df.hourly_consumption.astype('int32')\n",
    "train_df.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Train RCF Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "s3_train_path=os.path.join('s3://',data_bucket,'databrew_rcf','training','train.csv')\n",
    "s3_model_path=os.path.join('s3://',data_bucket,'databrew_rcf','model')\n",
    "\n",
    "wr.s3.to_csv(df=train_df,path=s3_train_path,header=False,index=False)\n",
    "training_channel=sm.inputs.TrainingInput(s3_data=s3_train_path,content_type='text/csv;label_size=0',distribution='ShardedByS3Key')\n",
    "channels={'train':training_channel}\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "rcf_algo_uri=image_uris.retrieve('randomcutforest',session.boto_region_name)\n",
    "rcf_estimator= sm.estimator.Estimator(rcf_algo_uri,role=role_arn,instance_count=1,instance_type='ml.m5.large',output_path=s3_model_path)\n",
    "rcf_estimator.set_hyperparameters(feature_dim=1)\n",
    "rcf_estimator.fit(channels)\n",
    "                            "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Deploy the trained model "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "rcf_endpoint_name='databrew-rcf-demo-endpoint'\n",
    "rcf_predictor=rcf_estimator.deploy(endpoint_name=rcf_endpoint_name,instance_type='ml.t2.medium',initial_instance_count=1,serializer=serializers.CSVSerializer(),deserializer=deserializers.JSONDeserializer())\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Predictions and Cleanup"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from statistics import mean,stdev\n",
    "test_df=pc_processed_df[(pc_processed_df['client_id']=='MT_012') & (pc_processed_df['timestamp'] >= '2014-01-01') &(pc_processed_df['hourly_consumption'] != 0)]\n",
    "test_df=test_df.tail(500)\n",
    "test_df_values=test_df['hourly_consumption'].astype('str').tolist()\n",
    "response=rcf_predictor.predict(test_df_values)\n",
    "scores = [datum[\"score\"] for datum in response[\"scores\"]]\n",
    "scores_mean=mean(scores)\n",
    "scores_std=stdev(scores)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### plot the prediction scores taking mean + or - 2*standard_deviation as the baseline  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "test_df['hourly_consumption'].plot(figsize=(40,10))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "plt.figure(figsize=(40,10))\n",
    "plt.plot(scores)\n",
    "plt.autoscale(tight=True)\n",
    "plt.axhline(y=scores_mean+2*scores_std,color='red')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Clean up by deleting the endpoint"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "rcf_predictor.delete_endpoint()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "conda_python3",
   "language": "python",
   "name": "conda_python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}