learning/SentencePredictions.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "import train_type as tp\n",
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "DATA_DIR = \"/mnt/big_drive/deeptype/data/\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "# function to get probabilities given a sentence split\n",
    "def get_prob(tagger_ins,sentence_splits):\n",
    "    ps = tagger_ins.predict_proba_sentences([sentence_splits])\n",
    "    output = [i for i in ps]\n",
    "    probs = output[0]['type']\n",
    "    return probs[0]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Load all tags\n",
    "(if you have multiple classifiers, read each classes.txt separetely and append to tags)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['aaa_no_instance_subclass_or_link',\n",
       " 'aaa_wikidata_prop',\n",
       " 'aaa_wikimedia_category_page',\n",
       " 'aaa_wikipedia_disambiguation',\n",
       " 'aaa_wikipedia_list',\n",
       " 'aaa_wikipedia_project_page',\n",
       " 'aaa_wikipedia_template_namespace',\n",
       " 'aaa_wikipedia_user_language_template',\n",
       " 'activity',\n",
       " 'aircraft',\n",
       " 'airport',\n",
       " 'algorithm',\n",
       " 'alphabet',\n",
       " 'anatomical_structure',\n",
       " 'astronomical_object',\n",
       " 'audio_visual_work',\n",
       " 'award',\n",
       " 'award_ceremony',\n",
       " 'battle',\n",
       " 'book_magazine_article',\n",
       " 'brand',\n",
       " 'bridge',\n",
       " 'character',\n",
       " 'chemical_compound',\n",
       " 'clothing',\n",
       " 'color',\n",
       " 'concept',\n",
       " 'country',\n",
       " 'crime',\n",
       " 'currency',\n",
       " 'data_format',\n",
       " 'date',\n",
       " 'developmental_biology_period',\n",
       " 'disease',\n",
       " 'electromagnetic_wave',\n",
       " 'event',\n",
       " 'facility',\n",
       " 'family',\n",
       " 'fictional_character',\n",
       " 'food',\n",
       " 'gas',\n",
       " 'gene',\n",
       " 'genre',\n",
       " 'geographical_object',\n",
       " 'geometric_shape',\n",
       " 'hazard',\n",
       " 'human',\n",
       " 'human_female',\n",
       " 'human_male',\n",
       " 'international_relations',\n",
       " 'kinship',\n",
       " 'lake',\n",
       " 'language',\n",
       " 'law',\n",
       " 'legal_action',\n",
       " 'legal_case',\n",
       " 'legislative_term',\n",
       " 'mathematical_object',\n",
       " 'mind',\n",
       " 'molecule',\n",
       " 'monument',\n",
       " 'mountain',\n",
       " 'musical_work',\n",
       " 'name',\n",
       " 'natural_phenomenon',\n",
       " 'number',\n",
       " 'organization',\n",
       " 'other_art_work',\n",
       " 'people',\n",
       " 'person_role',\n",
       " 'physical_object',\n",
       " 'physical_quantity',\n",
       " 'plant',\n",
       " 'populated_place',\n",
       " 'position',\n",
       " 'postal_code',\n",
       " 'radio_program',\n",
       " 'railroad',\n",
       " 'record_chart',\n",
       " 'region',\n",
       " 'religion',\n",
       " 'research',\n",
       " 'river',\n",
       " 'road_vehicle',\n",
       " 'sea',\n",
       " 'sex_toy',\n",
       " 'sexual_orientation',\n",
       " 'software',\n",
       " 'song',\n",
       " 'speech',\n",
       " 'sport',\n",
       " 'sport_event',\n",
       " 'sports_terminology',\n",
       " 'strategy',\n",
       " 'taxon',\n",
       " 'taxonomic_rank',\n",
       " 'title',\n",
       " 'train_station',\n",
       " 'union',\n",
       " 'unit_of_mass',\n",
       " 'value',\n",
       " 'vehicle',\n",
       " 'vehicle_brand',\n",
       " 'volcano',\n",
       " 'war',\n",
       " 'watercraft',\n",
       " 'weapon',\n",
       " 'website',\n",
       " 'other']"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# load all tags (if you have multiple classifiers read each classes.txt separetely and append to tags)\n",
    "tags = open('{}/classifications/type_classification/classes.txt'.format(DATA_DIR)).readlines()\n",
    "tags = [t.replace(\"\\n\",\"\") for t in tags]\n",
    "tags"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Load the LSTM Model \n",
    "Make sure you have sufficient memory in GPU, if you are using GPU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "INFO:tensorflow:Restoring parameters from /mnt/big_drive/deeptype/my_great_model_bckp/model.ckpt\n"
     ]
    }
   ],
   "source": [
    "tagger = tp.SequenceTagger('/mnt/big_drive/deeptype/my_great_model_bckp/')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Get probabilities for a sentence \n",
    "outputs probability vector for each word in a sentence, of size N*tags, \n",
    "where N is the number of words in the sentence and tags is the number of tags"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(7, 109)"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sentence = \"Man saw a jaguar in the jungle\"\n",
    "sent_splits = sentence.split()\n",
    "probs = get_prob(tagger,sent_splits)\n",
    "probs.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Probability for the word Man\n",
      "[('human_male', 0.14398718), ('mathematical_object', 0.34086773), ('other', 0.12715971)]\n",
      "\n",
      "\n",
      "Probability for the word saw\n",
      "[('mathematical_object', 0.16467695), ('physical_object', 0.2215165), ('other', 0.248256)]\n",
      "\n",
      "\n",
      "Probability for the word a\n",
      "[('mathematical_object', 0.13781491), ('physical_object', 0.10018583), ('taxon', 0.27661625), ('other', 0.2516701)]\n",
      "\n",
      "\n",
      "Probability for the word jaguar\n",
      "[('taxon', 0.6395821), ('other', 0.10057912)]\n",
      "\n",
      "\n",
      "Probability for the word in\n",
      "[('other', 0.44680375)]\n",
      "\n",
      "\n",
      "Probability for the word the\n",
      "[('region', 0.12901735), ('other', 0.46021998)]\n",
      "\n",
      "\n",
      "Probability for the word jungle\n",
      "[('region', 0.6417663), ('other', 0.13294013)]\n",
      "\n",
      "\n"
     ]
    }
   ],
   "source": [
    "threshold = 0.1 \n",
    "for k,p in enumerate(probs):\n",
    "    print(\"Types for the word {}\".format(sent_splits[k]))\n",
    "    print([(tags[k],p[k]) for k in np.where(p>threshold)[0]])\n",
    "    print(\"\\n\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}