Tensorflow in SWAP

Tensorflow_on_SWAP.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Install python dependencies\n",
    "===\n",
    "And package up a zip file to send to executors."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install tensorflow tensorflow_hub"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!cd venv; zip -qur ../spark_venv.zip ."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Download ELMo\n",
    "==="
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "# Make sure tensorflow downloads to a place we expect\n",
    "os.environ['TFHUB_CACHE_DIR'] = os.path.join(os.getcwd(), 'tf_hub')\n",
    "\n",
    "import tensorflow_hub as hub\n",
    "import tensorflow as tf\n",
    "\n",
    "session = tf.Session()\n",
    "embed = hub.Module(\"https://tfhub.dev/google/elmo/2\")\n",
    "session.run(tf.global_variables_initializer())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Quick verification embedding works locally"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "embeddings = embed([\n",
    "\"The quick brown fox jumps over the lazy dog.\",\n",
    "\"I am a sentence for which I would like to get its embedding\"])\n",
    "\n",
    "print(session.run(embeddings))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Bundle ELMo up to send to executors as well"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!cd tf_hub; zip -qur ../tf_hub.zip ."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Start up spark\n",
    "==="
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import findspark\n",
    "findspark.init('/usr/lib/spark2')\n",
    "from pyspark.ml.linalg import Vectors, VectorUDT\n",
    "from pyspark.sql import SparkSession, functions as F, types as T\n",
    "import os\n",
    "\n",
    "os.environ['PYSPARK_SUBMIT_ARGS'] = '--archives spark_venv.zip#venv,tf_hub.zip#tf_hub pyspark-shell'\n",
    "os.environ['PYSPARK_PYTHON'] = 'venv/bin/python'\n",
    "\n",
    "spark = (\n",
    "    SparkSession.builder\n",
    "    .appName('tensorflow: ELMo embedding')\n",
    "    .master('yarn')\n",
    "    .config('spark.executor.memoryOverhead', '2g')\n",
    "    .config('spark.dynamicAllocation.maxExecutors', 50)\n",
    "    .getOrCreate()\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Function to perform embedding on remote executors\n",
    "==="
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def batch(iterable, n):\n",
    "    batch = []\n",
    "    for x in iterable:\n",
    "        batch.append(x)\n",
    "        if len(x) == n:\n",
    "            yield batch\n",
    "            batch = []\n",
    "    if batch:\n",
    "        yield batch\n",
    "\n",
    "def embed_partition(rows, col):\n",
    "    import tensorflow as tf\n",
    "    import tensorflow_hub as hub\n",
    "    os.environ['TFHUB_CACHE_DIR'] = os.path.join(os.getcwd(), 'tf_hub')\n",
    "    sentences = [row[col] for row in rows]\n",
    "    with tf.Session() as session:\n",
    "        embed = hub.Module(\"https://tfhub.dev/google/elmo/2\")\n",
    "        session.run(tf.global_variables_initializer())\n",
    "        for batch_sentences in batch(sentences, 100):\n",
    "            embeddings = session.run(embed(batch_sentences))\n",
    "            for sentence, embedding in zip(batch_sentences, embeddings):\n",
    "                yield sentence, Vectors.dense(embedding)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Read some sentences and run them\n",
    "==="
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from functools import partial\n",
    "\n",
    "rdd_embedding = (\n",
    "    spark.table('ebernhardson.cirrus2hive')\n",
    "    .where(F.col('wikiid') == 'enwiki')\n",
    "    .where(F.col('dump_date') == '20190121')\n",
    "    .sample(withReplacement=False, fraction=0.00001)\n",
    "    .select(F.explode(F.col('heading')).alias('heading'))\n",
    "    .rdd.mapPartitions(partial(embed_partition, col='heading'))\n",
    ")\n",
    "\n",
    "df_embedding = (\n",
    "    spark.createDataFrame(rdd_embedding, T.StructType([\n",
    "        T.StructField('heading', T.StringType(), nullable=False),\n",
    "        T.StructField('embedding', VectorUDT(), nullable=False),\n",
    "    ]))\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "df_embedding.show()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.5.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}