notebooks/models/CRF.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "CRF trainer using the sklearn_crfsuite package (Python wrapper for CRFSuite): https://sklearn-crfsuite.readthedocs.io/en/latest/"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from collections import Counter\n",
    "\n",
    "import sklearn_crfsuite\n",
    "from sklearn_crfsuite import metrics\n",
    "\n",
    "from presidio_evaluator import InputSample\n",
    "from presidio_evaluator.models.crf_model import CRFModel\n",
    "from presidio_evaluator.data_generator import read_synth_dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "DATA_DATE = \"November 12 2019\""
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Source a dataset to use for training / testing:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "train_samples = read_synth_dataset(\"../../data/generated_train_{}.json\".format(DATA_DATE))\n",
    "test_samples = read_synth_dataset(\"../../data/generated_test_{}.json\".format(DATA_DATE))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [],
   "source": [
    "train_tagged = [sample for sample in train_samples if len(sample.spans) > 0]\n",
    "print(\"Kept {} train samples after removal of non-tagged samples\".format(len(train_tagged)))\n",
    "train_data = InputSample.create_conll_dataset(train_tagged)\n",
    "\n",
    "test_data = InputSample.create_conll_dataset(test_samples)\n",
    "test_data.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Turn every sentence into a list of lists (list of tokens + pos + label)\n",
    "test_sents=test_data.groupby('sentence')[['text','pos','label']].apply(lambda x: x.values.tolist())\n",
    "train_sents=train_data.groupby('sentence')[['text','pos','label']].apply(lambda x: x.values.tolist())\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Create features for CRF"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "CRFModel.sent2features(train_sents[0])[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%%time\n",
    "X_train = [CRFModel.sent2features(s) for s in train_sents]\n",
    "y_train = [CRFModel.sent2labels(s) for s in train_sents]\n",
    "\n",
    "X_test = [CRFModel.sent2features(s) for s in test_sents]\n",
    "y_test = [CRFModel.sent2labels(s) for s in test_sents]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "%%time\n",
    "crf = sklearn_crfsuite.CRF(\n",
    "    algorithm='lbfgs',\n",
    "    c1=0.1,\n",
    "    c2=0.1,\n",
    "    max_iterations=100,\n",
    "    all_possible_transitions=True\n",
    ")\n",
    "crf.fit(X_train, y_train)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Save trained model to pickle"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pickle\n",
    "with open(\"../../models/crf.pickle\",'wb') as f:\n",
    "    pickle.dump(crf, f,protocol=pickle.HIGHEST_PROTOCOL)\n",
    "    "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Open saved model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "with open(\"../../models/crf.pickle\", 'rb') as f:\n",
    "    crf = pickle.load(f)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Extract info and predictions from model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "labels = list(crf.classes_)\n",
    "labels.remove('O')\n",
    "labels"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "y_pred = crf.predict(X_test)\n",
    "metrics.flat_f1_score(y_test, y_pred,\n",
    "                      average='weighted', labels=labels)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "## predict one:\n",
    "y_5_pred = crf.predict([X_test[5]])\n",
    "y_5_pred[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# group B and I results\n",
    "sorted_labels = sorted(\n",
    "    labels,\n",
    "    key=lambda name: (name[1:], name[0])\n",
    ")\n",
    "print(metrics.flat_classification_report(\n",
    "    y_test, y_pred, labels=sorted_labels, digits=3\n",
    "))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Model explainability"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_transitions(trans_features):\n",
    "    for (label_from, label_to), weight in trans_features:\n",
    "        print(\"%-6s -> %-7s %0.6f\" % (label_from, label_to, weight))\n",
    "\n",
    "print(\"Top likely transitions:\")\n",
    "print_transitions(Counter(crf.transition_features_).most_common(20))\n",
    "\n",
    "print(\"\\nTop unlikely transitions:\")\n",
    "print_transitions(Counter(crf.transition_features_).most_common()[-20:])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_state_features(state_features):\n",
    "    for (attr, label), weight in state_features:\n",
    "        print(\"%0.6f %-8s %s\" % (weight, label, attr))\n",
    "\n",
    "print(\"Top positive:\")\n",
    "print_state_features(Counter(crf.state_features_).most_common(30))\n",
    "\n",
    "print(\"\\nTop negative:\")\n",
    "print_state_features(Counter(crf.state_features_).most_common()[-30:])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "PyCharm (presidio-research)",
   "language": "python",
   "name": "pycharm-c8930cf3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.8.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}