serving-input-reception-bug

serving-input-reception-bug.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Mishandling of estimators with single, unnamed feature\n",
    "\n",
    "The [tf.estimator.Estimator](https://www.tensorflow.org/api_docs/python/tf/estimator/Estimator) interface allows users to provide a `model_fn` which accepts features either within a single tensor or within a dictionary mapping strings to tensors.\n",
    "\n",
    "The Estimator `export_savedmodel` method requires a `serving_input_receiver_fn` argument, which is a function of no arguments that produces a [ServingInputReceiver](https://www.tensorflow.org/api_docs/python/tf/estimator/export/ServingInputReceiver). The features tensors from this `ServingInputReceiver` are passed to the `model_fn` for serving.\n",
    "\n",
    "Upon instantiation, the `ServingInputReceiver` wraps single tensor features into a dictionary. This raises an error for estimators whose `model_fn` expects a single tensor as its `features` argument."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Sample estimator"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import tensorflow as tf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def model_fn(features, labels, mode):    \n",
    "    output = tf.multiply(features, features)\n",
    "    \n",
    "    prediction_output = tf.estimator.export.PredictOutput({'trololol': output})\n",
    "    \n",
    "    return tf.estimator.EstimatorSpec(\n",
    "        mode=mode,\n",
    "        predictions=features,\n",
    "        train_op=tf.no_op(),\n",
    "        loss=tf.constant(1, dtype=tf.float32),\n",
    "        export_outputs={tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: prediction_output}\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "MODEL_DIR = '/tmp/dummy-model'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "dummy = tf.estimator.Estimator(model_fn, model_dir=MODEL_DIR)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dummy.train(lambda: (tf.constant([1, 2, 3, 4, 5], dtype=tf.float32), None), steps=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Export woes\n",
    "\n",
    "The following `serving_input_receiver_fn` uses `ServingInputReceiver`;"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def serving_input_receiver_fn():\n",
    "    feature_tensor = tf.placeholder(tf.float32, [None, 1])\n",
    "    return tf.estimator.export.ServingInputReceiver(feature_tensor, {'input': feature_tensor})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "BASE_EXPORT_DIR = '/tmp/dummy-model/export'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dummy.export_savedmodel(BASE_EXPORT_DIR, serving_input_receiver_fn)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "As you can see, there is no mechanism in the `model_fn` to handle this dictionary, and we can't expect there to be one because the user provides the implementation."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Fix\n",
    "\n",
    "Either we should change the estimator interface to only accept a `model_fn` which takes its features in a dictionary, OR we should provide a different type of `ServingInputReceiver`. For example, something like this:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "class FlatServingInputReceiver(object):\n",
    "    def __init__(self, feature):\n",
    "        self.features = feature\n",
    "        self.receiver_tensors = {'input': feature}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def flat_serving_input_receiver_fn():\n",
    "    feature_tensor = tf.placeholder(tf.float32, [None, 1])\n",
    "    return FlatServingInputReceiver(feature_tensor)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dummy.export_savedmodel(BASE_EXPORT_DIR, flat_serving_input_receiver_fn)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}