notebook_examples/feature_store_embeddings.ipynb

{
 "cells": [
  {
   "cell_type": "raw",
   "id": "16cafe86",
   "metadata": {
    "vscode": {
     "languageId": "raw"
    }
   },
   "source": [
    "@notebook{feature_store_embeddings.ipynb,\n",
    "    title: Using feature store for storage, retrieval, versioning and time travel of embeddings,\n",
    "    summary: Feature store to store embeddings, version embeddings and time travel of embeddings.,\n",
    "    developed_on: fspyspark32_p38_cpu_v3,\n",
    "    keywords: feature store, llm, embeddings,\n",
    "    license: Universal Permissive License v 1.0\n",
    "}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "33e7cf60",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-05-24T08:26:08.572567Z",
     "start_time": "2023-05-24T08:26:08.328013Z"
    }
   },
   "outputs": [],
   "source": [
    "!odsc conda install -s fspyspark32_p38_cpu_v3"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ba5e4106",
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install tensorflow"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f3e34844",
   "metadata": {},
   "source": [
    "Oracle Data Science service sample notebook.\n",
    "\n",
    "Copyright (c) 2022 Oracle, Inc. All rights reserved. Licensed under the [Universal Permissive License v 1.0](https://oss.oracle.com/licenses/upl).\n",
    "\n",
    "***\n",
    "\n",
    "# <font color=\"red\">Embeddings in feature store</font>\n",
    "<p style=\"margin-left:10%; margin-right:10%;\">by the <font color=\"teal\">Oracle Cloud Infrastructure Data Science Service.</font></p>\n",
    "\n",
    "---\n",
    "# Overview:\n",
    "---\n",
    "Managing many datasets, data-sources and transformations for machine learning is complex and costly. Poorly cleaned data, data issues, bugs in transformations, data drift and training serving skew all leads to increased model development time and worse model performance. Here, feature store is well positioned to solve many of the problems since it provides a centralised way to transform and access data for training and serving time and helps defines a standardised pipeline for ingestion of data and querying of data. This notebook demonstrates how to use feature store within a long lasting [Oracle Cloud Infrastructure Data Flow](https://docs.oracle.com/en-us/iaas/data-flow/using/home.htm) cluster.\n",
    "\n",
    "* **Storage of Pre-trained Embeddings**: One of the primary functions of an embedding feature store is to store pre-trained word embeddings, such as Word2Vec, GloVe, FastText, or BERT embeddings. These embeddings are learned from massive text corpora and contain information about word semantics, which can be valuable for various NLP tasks like text classification, named entity recognition, sentiment analysis, and more.\n",
    "\n",
    "* **Customized Embeddings**: Some organizations prefer to train their embeddings on domain-specific data or fine-tune pre-trained embeddings to better suit their particular use cases. An embedding feature store can store these custom embeddings for reuse across different projects or models. User can store custom embeddings using ``Transformation`` construct in feature store.\n",
    "\n",
    "* **Efficient Retrieval**: Embedding feature stores are optimized for fast and efficient retrieval of embeddings. This is crucial because embeddings can be high-dimensional and computationally expensive to calculate. By storing them in a dedicated store, you can avoid the need to recalculate embeddings for the same data repeatedly.\n",
    "\n",
    "* **Consistency and Versioning**: An embedding feature store often includes version control and management capabilities. This ensures that models use consistent embeddings even as new versions of pre-trained embeddings become available. It helps maintain reproducibility and ensures that model behavior remains stable over time.\n",
    "\n",
    "* **Scaling for Large Datasets**: For organizations dealing with massive datasets, efficiently storing and retrieving embeddings is a significant challenge. Feature store using spark as execution engine can be used to handle large-scale embeddings and provides mechanisms for distributed storage and retrieval.\n",
    "\n",
    "<div>\n",
    "    <img src=\"https://objectstorage.us-ashburn-1.oraclecloud.com/p/hh2NOgFJbVSg4amcLM3G3hkTuHyBD-8aE_iCsuZKEvIav1Wlld-3zfCawG4ycQGN/n/ociodscdev/b/oci-feature-store/o/beta/_images/feature-store-overview.png\" width=\"1350\" height=\"700\" />\n",
    "</div>\n",
    "\n",
    "\n",
    "Compatible conda pack: [PySpark 3.2 and Feature store](https://docs.oracle.com/iaas/data-science/using/conda-pyspark-fam.htm) for CPU on Python 3.8\n",
    "\n",
    "## Contents:\n",
    "\n",
    "- <a href=\"#concepts\">1. Introduction</a>\n",
    "- <a href='#pre-requisites'>1. Pre-requisites</a>\n",
    "    - <a href='#policies'>2.1 Policies</a>\n",
    "    - <a href='#prerequisites_authentication'>2.2 Authentication</a>\n",
    "    - <a href='#prerequisites_variables'>2.3 Variables</a>\n",
    "- <a href='#featurestore_querying'>3. Feature store querying</a>\n",
    "    - <a href='#data_exploration'>3.1. Exploration of data in feature store</a>\n",
    "    - <a href='#load_featuregroup'>3.2. Load feature groups</a>\n",
    "    - <a href='#explore_featuregroup'>3.3. Explore feature groups</a>\n",
    "    - <a href='#select_subset_featuregroup'>3.4. Select subset of features</a>\n",
    "    - <a href='#filter_featuregroup'>3.5. Filter feature groups</a>\n",
    "    - <a href='#join_featuregroup'>3.6. Apply joins on feature group</a>\n",
    "    - <a href='#create_dataset'>3.7. Create dataset from multiple or one feature group</a>\n",
    "    - <a href='#sql_query'>3.8 Free form sql query</a>\n",
    "- <a href='#ref'>4. References</a>\n",
    "\n",
    "---\n",
    "\n",
    "**Important:**\n",
    "\n",
    "Placeholder text for required values are surrounded by angle brackets that must be removed when adding the indicated content. For example, when adding a database name to `database_name = \"<database_name>\"` would become `database_name = \"production\"`.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9ca2b422",
   "metadata": {},
   "source": [
    "<a id=\"concepts\"></a>\n",
    "# 1. Introduction\n",
    "\n",
    "Oracle feature store is a stack based solution that is deployed in the customer enclave using OCI resource manager. Customer can stand up the service with infrastructure in their own tenancy. The service consists of API which are deployed in customer tenancy using resource manager.\n",
    "\n",
    "The following are some key terms that will help you understand OCI Data Science Feature Store:\n",
    "\n",
    "\n",
    "* **Feature Vector**: Set of feature values for any one primary/identifier key. Eg. All/subset of features of customer id ‘2536’ can be called as one feature vector.\n",
    "\n",
    "* **Feature**: A feature is an individual measurable property or characteristic of a phenomenon being observed.\n",
    "\n",
    "* **Entity**: An entity is a group of semantically related features. The first step a consumer of features would typically do when accessing the feature store service is to list the entities and the entities associated features. Another way to look at it is that an entity is an object or concept that is described by its features. Examples of entities could be customer, product, transaction, review, image, document, etc.\n",
    "\n",
    "* **Feature Group**: A feature group in a feature store is a collection of related features that are often used together in ml models. It serves as an organizational unit within the feature store for users to manage, version and share features across different ml projects. By organizing features into groups, data scientists and ml engineers can efficiently discover, reuse and collaborate on features reducing the redundant work and ensuring consistency in feature engineering.\n",
    "\n",
    "* **Feature Group Job**: Feature group job is the execution instance of a feature group. Each feature group job will include validation results and statistics results.\n",
    "\n",
    "* **Dataset**: A dataset is a collection of feature that are used together to either train a model or perform model inference.\n",
    "\n",
    "* **Dataset Job**: Dataset job is the execution instance of a dataset. Each dataset job will include validation results and statistics results."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "310cc3a2",
   "metadata": {},
   "source": [
    "<a id='pre-requisites'></a>\n",
    "# 2. Pre-requisites\n",
    "\n",
    "Data Flow Sessions are accessible through the following conda environment:\n",
    "\n",
    "* **PySpark 3.2, Feature store 1.0 and Data Flow 1.0 (fs_pyspark32_p38_cpu_v1)**\n",
    "\n",
    "The [Data Catalog Hive Metastore](https://docs.oracle.com/en-us/iaas/data-catalog/using/metastore.htm) provides schema definitions for objects in structured and unstructured data assets. The Metastore is the central metadata repository to understand tables backed by files on object storage. You can customize `fs_pyspark32_p38_cpu_v1`, publish it, and use it as a runtime environment for a Data Flow session cluster. The metastore id of hive metastore is tied to feature store construct of feature store service.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0d46f58e",
   "metadata": {},
   "source": [
    "<a id='setup_spark-defaults'></a>\n",
    "### `spark-defaults.conf`\n",
    "\n",
    "The `spark-defaults.conf` file is used to define the properties that are used by Spark. A templated version is installed when you install a Data Science conda environment that supports PySpark. However, you must update the template so that the Data Catalog metastore can be accessed. You can do this manually. However, the `odsc data-catalog config` commandline tool is ideal for setting up the file because it gathers information about your environment, and uses that to build the file.\n",
    "\n",
    "The `odsc data-catalog config` command line tool needs the `--metastore` option to define the Data Catalog metastore OCID. No other command line option is needed because settings have default values, or they take values from your notebook session environment. Following are common parameters that you may need to override.\n",
    "\n",
    "The `--authentication` option sets the authentication mode. It supports resource principal and API keys. The preferred method for authentication is resource principal, which is sent with `--authentication resource_principal`. If you want to use API keys, then use the `--authentication api_key` option. If the `--authentication` isn't specified, API keys are used. When API keys are used, information from the OCI configuration file is used to create the `spark-defaults.conf` file.\n",
    "\n",
    "Object Storage and Data Catalog are regional services. By default, the region is set to the region your notebook session is running in. This information is taken from the environment variable, `NB_REGION`. Use the `--region` option to override this behavior.\n",
    "\n",
    "The default location of the `spark-defaults.conf` file is `/home/datascience/spark_conf_dir` as defined in the `SPARK_CONF_DIR` environment variable. Use the `--output` option to define the directory where to write the file.\n",
    "\n",
    "You need to determine what settings are appropriate for your configuration. However, the following works for most configurations and is run in a terminal window.\n",
    "\n",
    "```bash\n",
    "odsc data-catalog config --authentication resource_principal --metastore <metastore_id>\n",
    "```\n",
    "For more assistance, use the following command in a terminal window:\n",
    "\n",
    "```bash\n",
    "odsc data-catalog config --help\n",
    "```\n",
    "\n",
    "<a id='setup_session'></a>\n",
    "### Session Setup\n",
    "\n",
    "The notebook makes connections to the Data Catalog metastore and Object Storage. In the next cell, specify the bucket URI to act as the data warehouse. Use the `warehouse_uri` variable with the `oci://<bucket_name>@<namespace_name>/<key>` format. Update the variable `metastore_id` with the OCID of the Data Catalog metastore."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0a4615d2",
   "metadata": {},
   "source": [
    "python -m pip install --pre oracle-ads<a id='policies'></a>\n",
    "### 2.1. Policies\n",
    "This section covers the creation of dynamic groups and policies needed to use the service.\n",
    "\n",
    "* [Data Flow Policies](https://docs.oracle.com/iaas/data-flow/using/policies.htm/)\n",
    "* [Data Catalog Metastore Required Policies](https://docs.oracle.com/en-us/iaas/data-catalog/using/metastore.htm)\n",
    "* [Getting Started with Data Flow](https://docs.oracle.com/iaas/data-flow/using/dfs_getting_started.htm)\n",
    "* [About Data Science Policies](https://docs.oracle.com/iaas/data-science/using/policies.htm)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0b8a6ab0",
   "metadata": {},
   "source": [
    "<a id=\"prerequisites_authentication\"></a>\n",
    "### 2.2. Authentication\n",
    "The [Oracle Accelerated Data Science SDK (ADS)](https://docs.oracle.com/iaas/tools/ads-sdk/latest/index.html) controls the authentication mechanism with the notebook cluster.<br>\n",
    "To setup authentication use the ```ads.set_auth(\"resource_principal\")``` or ```ads.set_auth(\"api_key\")```."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8e49e7ee",
   "metadata": {
    "ExecuteTime": {
     "start_time": "2023-05-24T08:26:08.577504Z"
    },
    "is_executing": true,
    "pycharm": {
     "is_executing": true
    }
   },
   "outputs": [],
   "source": [
    "import ads\n",
    "ads.set_auth(auth=\"resource_principal\", client_kwargs={\"fs_service_endpoint\": \"https://{api_gateway}/20230101\"})"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "df8f99a8",
   "metadata": {},
   "source": [
    "<a id=\"prerequisites_variables\"></a>\n",
    "### 2.3. Variables\n",
    "To run this notebook, you must provide some information about your tenancy configuration. To create and run a feature store, you must specify a `<compartment_id>` and bucket `<metastore_id>` for offline feature store."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "946db756",
   "metadata": {
    "pycharm": {
     "is_executing": true
    }
   },
   "outputs": [],
   "source": [
    "import os\n",
    "\n",
    "compartment_id = os.environ.get(\"NB_SESSION_COMPARTMENT_OCID\")\n",
    "metastore_id = \"<metastore_id>\""
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3d342464",
   "metadata": {},
   "source": [
    "<a id=\"featurestore_querying\"></a>\n",
    "# 3. Feature group querying\n",
    "By default the **PySpark 3.2, Feature store and Data Flow** conda environment includes pre-installed [great-expectations](https://legacy.docs.greatexpectations.io/en/latest/reference/core_concepts/validation.html) and [deeque](https://github.com/awslabs/deequ) libraries. The joining functionality is heavily inspired by the APIs used by Pandas to merge, join or filter DataFrames. The APIs allow you to specify which features to select from which feature group, how to join them and which features to use in join conditions."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e8d1d24f",
   "metadata": {
    "pycharm": {
     "is_executing": true
    }
   },
   "outputs": [],
   "source": [
    "import warnings\n",
    "warnings.filterwarnings(\"ignore\", message=\"iteritems is deprecated\")\n",
    "warnings.filterwarnings(\"ignore\", category=DeprecationWarning)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a4308922",
   "metadata": {
    "pycharm": {
     "is_executing": true
    }
   },
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "from ads.feature_store.feature_store import FeatureStore\n",
    "from ads.feature_store.feature_group import FeatureGroup\n",
    "from ads.feature_store.model_details import ModelDetails\n",
    "from ads.feature_store.dataset import Dataset\n",
    "from ads.feature_store.common.enums import DatasetIngestionMode\n",
    "\n",
    "\n",
    "from ads.feature_store.feature_group_expectation import ExpectationType\n",
    "from great_expectations.core import ExpectationSuite, ExpectationConfiguration\n",
    "from ads.feature_store.feature_store_registrar import FeatureStoreRegistrar"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "53243e06",
   "metadata": {},
   "source": [
    "<a id=\"data_exploration\"></a>\n",
    "### 3.1. Exploration of data in feature store"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c5404e1e",
   "metadata": {},
   "source": [
    "#### 3.1.1 Load SQuAD dataset\n",
    "\n",
    "There are several important fields here:\n",
    "\n",
    "* answers: the starting location of the answer token and the answer text.\n",
    "* context: background information from which the model needs to extract the answer.\n",
    "* question: the question a model should answer."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "bd3a6ad5",
   "metadata": {},
   "outputs": [],
   "source": [
    "from datasets import load_dataset\n",
    "\n",
    "squad = load_dataset(\"squad\", split=\"train[:5000]\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "547ba470",
   "metadata": {},
   "source": [
    "<a id=\"load_featuregroup\"></a>\n",
    "### 3.2. Create feature store logical entities"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7702d75d",
   "metadata": {},
   "source": [
    "#### 3.2.1 Feature Store\n",
    "Feature store is the top level entity for feature store service"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "96a73af4",
   "metadata": {
    "pycharm": {
     "is_executing": true
    }
   },
   "outputs": [],
   "source": [
    "feature_store_resource = (\n",
    "    FeatureStore().\n",
    "    with_description(\"SQUAD Dataset Feature Store\").\n",
    "    with_compartment_id(compartment_id).\n",
    "    with_name(\"SQUAD details\").\n",
    "    with_offline_config(metastore_id=metastore_id)\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5bdd997f",
   "metadata": {},
   "source": [
    "<a id=\"create_feature_store\"></a>\n",
    "##### Create Feature Store\n",
    "\n",
    "Call the ```.create()``` method of the Feature store instance to create a feature store."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "cdbe74a3",
   "metadata": {
    "pycharm": {
     "is_executing": true
    }
   },
   "outputs": [],
   "source": [
    "feature_store = feature_store_resource.create()\n",
    "feature_store"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3cdb7982",
   "metadata": {},
   "source": [
    "#### 3.2.2 Entity\n",
    "An entity is a logical segregation of feature store entities"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6e4b3ca1",
   "metadata": {},
   "outputs": [],
   "source": [
    "entity = feature_store.create_entity(\n",
    "    name=\"Squad Entity\",\n",
    "    description=\"description for Squad entity details\"\n",
    ")\n",
    "entity"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "62588a98",
   "metadata": {},
   "source": [
    "#### 3.2.3 Transformation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "89ae77ad",
   "metadata": {},
   "outputs": [],
   "source": [
    "def squad_embedding_transformation(df):\n",
    "    from datasets import Dataset\n",
    "    from transformers import AutoTokenizer\n",
    "    import json\n",
    "    import numpy as np\n",
    "\n",
    "    tokenizer = AutoTokenizer.from_pretrained(\"distilbert-base-uncased\")\n",
    "    \n",
    "    def preprocess_function(examples):\n",
    "        questions = [q.strip() for q in examples[\"question\"]]\n",
    "\n",
    "        inputs = tokenizer(\n",
    "            questions,\n",
    "            examples[\"context\"],\n",
    "            max_length=384,\n",
    "            truncation=\"only_second\",\n",
    "            return_offsets_mapping=True,\n",
    "            padding=\"max_length\",\n",
    "        )\n",
    "\n",
    "        offset_mapping = inputs.pop(\"offset_mapping\")\n",
    "        answers = examples[\"answers\"]\n",
    "        \n",
    "        start_positions = []\n",
    "        end_positions = []\n",
    "\n",
    "        for i, offset in enumerate(offset_mapping):\n",
    "            answer = answers[i]\n",
    "            start_char = answer[\"answer_start\"][0]\n",
    "            end_char = answer[\"answer_start\"][0] + len(answer[\"text\"][0])\n",
    "            sequence_ids = inputs.sequence_ids(i)\n",
    "\n",
    "            # Find the start and end of the context\n",
    "            idx = 0\n",
    "            while sequence_ids[idx] != 1:\n",
    "                idx += 1\n",
    "            context_start = idx\n",
    "            while sequence_ids[idx] == 1:\n",
    "                idx += 1\n",
    "            context_end = idx - 1\n",
    "\n",
    "            # If the answer is not fully inside the context, label it (0, 0)\n",
    "            if offset[context_start][0] > end_char or offset[context_end][1] < start_char:\n",
    "                start_positions.append(0)\n",
    "                end_positions.append(0)\n",
    "            else:\n",
    "                # Otherwise it's the start and end token positions\n",
    "                idx = context_start\n",
    "                while idx <= context_end and offset[idx][0] <= start_char:\n",
    "                    idx += 1\n",
    "                start_positions.append(idx - 1)\n",
    "\n",
    "                idx = context_end\n",
    "                while idx >= context_start and offset[idx][1] >= end_char:\n",
    "                    idx -= 1\n",
    "                end_positions.append(idx + 1)\n",
    "\n",
    "        inputs[\"start_positions\"] = start_positions\n",
    "        inputs[\"end_positions\"] = end_positions\n",
    "        return inputs\n",
    "\n",
    "    df['answers'] = df['answers'].apply(json.loads)\n",
    "\n",
    "    # Convert lists back to NumPy arrays within dictionaries\n",
    "    df['answers'] = df['answers'].apply(lambda x: {k: np.array(v) for k, v in x.items()})\n",
    "    dataset = Dataset.from_pandas(df)\n",
    "    dataset = dataset.train_test_split(test_size=0.2)\n",
    "    dataset = dataset.map(preprocess_function, batched=True, remove_columns=dataset[\"train\"].column_names)\n",
    "    \n",
    "    return dataset[\"train\"].to_pandas()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "97083c7c",
   "metadata": {},
   "outputs": [],
   "source": [
    "from ads.feature_store.transformation import TransformationMode\n",
    "\n",
    "squad_transformation = feature_store.create_transformation(\n",
    "    transformation_mode=TransformationMode.PANDAS,\n",
    "    source_code_func=squad_embedding_transformation,\n",
    "    name=\"squad_embedding_transformation\",\n",
    ")\n",
    "\n",
    "squad_transformation"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fa228b02",
   "metadata": {},
   "source": [
    "#### 3.2.3 Feature group\n",
    "A feature group is an object that represents a logical group of time-series feature data as it is found in a datasource."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c148e158",
   "metadata": {},
   "source": [
    "<a id=\"create_feature_group_flights\"></a>\n",
    "##### SQUad Dataset Embedding Feature Group\n",
    "\n",
    "Create feature group for SQUad Dataset Embedding\n",
    "\n",
    "<div>\n",
    "    <img src=\"https://objectstorage.us-ashburn-1.oraclecloud.com/p/hh2NOgFJbVSg4amcLM3G3hkTuHyBD-8aE_iCsuZKEvIav1Wlld-3zfCawG4ycQGN/n/ociodscdev/b/oci-feature-store/o/beta/_images/feature_group_flights.gif\" width=\"700\" height=\"350\" />\n",
    "</div>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "721ae60d",
   "metadata": {},
   "outputs": [],
   "source": [
    "import json\n",
    "squad_df = squad.to_pandas()\n",
    "\n",
    "# Convert NumPy arrays within dictionaries to lists\n",
    "squad_df['answers'] = squad_df['answers'].apply(lambda x: {k: v.tolist() for k, v in x.items()})\n",
    "\n",
    "# Convert the 'data' column back to dictionaries of NumPy ndarrays\n",
    "squad_df['answers'] = squad_df['answers'].apply(json.dumps)\n",
    "\n",
    "squad_embedding_feature_group = (\n",
    "    FeatureGroup()\n",
    "    .with_feature_store_id(feature_store.id)\n",
    "    .with_primary_keys([])\n",
    "    .with_name(\"squad_embedding_feature_group\")\n",
    "    .with_entity_id(entity.id)\n",
    "    .with_compartment_id(compartment_id)\n",
    "    .with_schema_details_from_dataframe(squad_df)\n",
    "    .with_transformation_id(squad_transformation.id)\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c6897c26",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.create()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "2f0374de",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ba9c27a9",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.materialise(squad_df)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d36c73ea",
   "metadata": {},
   "source": [
    "<a id=\"explore_featuregroup\"></a>\n",
    "### 3.3. Explore feature groups"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a050d793",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.get_statistics().to_viz()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f72a14cc",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.history().show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "67b4e0c1",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.as_of(version_number=0).show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ae448faf",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.get_features_df()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8893c22d",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.get_validation_output().to_pandas()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9296cd2c",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.select().show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d47e04ea",
   "metadata": {},
   "source": [
    "<a id=\"select_subset_featuregroup\"></a>\n",
    "### 3.4. Select subset of features"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9187e157",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.select(['input_ids', 'attention_mask']).show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1c07330c",
   "metadata": {},
   "source": [
    "<a id=\"filter_featuregroup\"></a>\n",
    "### 3.5. Filter feature groups"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9df4ae16",
   "metadata": {},
   "outputs": [],
   "source": [
    "squad_embedding_feature_group.filter(squad_embedding_feature_group.start_positions == \"106\").show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b5e0f508",
   "metadata": {},
   "outputs": [],
   "source": [
    "tranining_df = squad_embedding_feature_group.select().read().toPandas()\n",
    "tranining_df"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3bfa5f39",
   "metadata": {},
   "source": [
    "## 4. Traning Model using Bert Model\n",
    "\n",
    "You’re ready to start training your model now! Load DistilBERT with AutoModelForQuestionAnswering:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f079d8f8",
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import create_optimizer\n",
    "\n",
    "batch_size = 16\n",
    "num_epochs = 2\n",
    "total_train_steps = (len(tranining_df) // batch_size) * num_epochs\n",
    "optimizer, schedule = create_optimizer(\n",
    "    init_lr=2e-5,\n",
    "    num_warmup_steps=0,\n",
    "    num_train_steps=total_train_steps,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "64defca4",
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import TFAutoModelForQuestionAnswering\n",
    "\n",
    "model = TFAutoModelForQuestionAnswering.from_pretrained(\"distilbert-base-uncased\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3cce2847",
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import DefaultDataCollator\n",
    "\n",
    "data_collator = DefaultDataCollator(return_tensors=\"tf\")\n",
    "from datasets import Dataset\n",
    "tf_train_set = model.prepare_tf_dataset(\n",
    "    Dataset.from_pandas(tranining_df),\n",
    "    shuffle=True,\n",
    "    batch_size=16,\n",
    "    collate_fn=data_collator,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "33f5f78f",
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "\n",
    "model.compile(optimizer=optimizer)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "04339166",
   "metadata": {},
   "outputs": [],
   "source": [
    "model.fit(x=tf_train_set, epochs=3)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "186ef0d8",
   "metadata": {},
   "source": [
    "## 5. Evaluation\n",
    "\n",
    "Evaluation for question answering requires a significant amount of postprocessing. To avoid taking up too much of your time, this guide skips the evaluation step. The Trainer still calculates the evaluation loss during training so you’re not completely in the dark about your model’s performance."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "45551eff",
   "metadata": {},
   "outputs": [],
   "source": [
    "question = \"How many programming languages does BLOOM support?\"\n",
    "context = \"BLOOM has 176 billion parameters and can generate text in 46 languages natural languages and 13 programming languages.\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d18b79d7",
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import pipeline\n",
    "\n",
    "question_answerer = pipeline(\"question-answering\", model=\"my_awesome_qa_model\")\n",
    "question_answerer(question=question, context=context)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2b4f0ff5",
   "metadata": {},
   "source": [
    "<a id='ref'></a>\n",
    "# References\n",
    "\n",
    "- [ADS Library Documentation](https://accelerated-data-science.readthedocs.io/en/latest/index.html)\n",
    "- [Data Science YouTube Videos](https://www.youtube.com/playlist?list=PLKCk3OyNwIzv6CWMhvqSB_8MLJIZdO80L)\n",
    "- [OCI Data Science Documentation](https://docs.cloud.oracle.com/en-us/iaas/data-science/using/data-science.htm)\n",
    "- [Oracle Data & AI Blog](https://blogs.oracle.com/datascience/)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "99fe3d5b",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "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": "3.7.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}