2. evalutation.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# evalutation code after train \n",
    "190605\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "# coding=utf-8\n",
    "from datetime import datetime\n",
    "import torch\n",
    "import _pickle as cPickle\n",
    "from torch.autograd import Variable\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "\n",
    "from loader import *\n",
    "from utils import *"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{   'char_mode': 'CNN',\n",
      "    'crf': 0,\n",
      "    'dev': 'data/eng.testa',\n",
      "    'eval_path': 'evaluation',\n",
      "    'eval_script': 'evaluation/conlleval',\n",
      "    'eval_temp': 'evaluation/temp',\n",
      "    'loss': 'loss.txt',\n",
      "    'map_path': 'models/mapping.pkl',\n",
      "    'model_name': 'models/lstm_crf.model',\n",
      "    'models_path': 'models',\n",
      "    'name': 'lstm_crf.model',\n",
      "    'score': 'evaluation/temp/score.txt',\n",
      "    'test': 'data/eng.testb',\n",
      "    'train': 'data/eng.train',\n",
      "    'use_gpu': 1}\n"
     ]
    }
   ],
   "source": [
    "class Myconfig():\n",
    "    def __init__(self):\n",
    "        self.train        = 'data/eng.train'  # Train set location\n",
    "        self.dev          = 'data/eng.testa'  # Dev set location\n",
    "        self.test         = 'data/eng.testb'  # Test set location\n",
    "        \n",
    "        self.score        = 'evaluation/temp/score.txt'  # score file location\n",
    "        self.crf          = 0  # Use CRF (0 to disable)\n",
    "         \n",
    "        self.use_gpu      = 1  # whether or not to ues gpu\n",
    "        self.loss         = 'loss.txt'  # loss file location\n",
    "        self.map_path     = 'models/mapping.pkl'  # model path\n",
    "        self.char_mode    = 'CNN'  # ['CNN', 'LSTM']  # char_CNN or char_LSTM\n",
    "        self.models_path  = \"models\"\n",
    "        self.name         = 'lstm_crf.model'  # model name\n",
    "        self.model_name   = os.path.join(self.models_path, self.name)\n",
    "        self.eval_path    = \"evaluation\"\n",
    "        self.eval_temp    = os.path.join(self.eval_path, \"temp\")\n",
    "        self.eval_script  = os.path.join(self.eval_path, \"conlleval\")\n",
    "        \n",
    "opts = Myconfig()\n",
    "def _print_config(config):\n",
    "    import pprint\n",
    "    pp = pprint.PrettyPrinter(indent=4)\n",
    "    pp.pprint(vars(config))\n",
    "_print_config(opts)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "with open(opts.map_path, 'rb') as f:\n",
    "    mappings = cPickle.load(f)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "gpu possible!\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1080x216 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "word_to_id  = mappings['word_to_id']\n",
    "tag_to_id   = mappings['tag_to_id']\n",
    "id_to_tag   = {k[1]: k[0] for k in tag_to_id.items()}\n",
    "char_to_id  = mappings['char_to_id']\n",
    "parameters  = mappings['parameters']\n",
    "word_embeds = mappings['word_embeds']\n",
    "epoch_losses= mappings['epoch_losses']\n",
    "all_F       = mappings['all_F']\n",
    "\n",
    "lower = parameters.lower\n",
    "zeros = parameters.zeros\n",
    "tag_scheme = parameters.tag_scheme\n",
    "\n",
    "if opts.use_gpu == 1 and torch.cuda.is_available():  print('gpu possible!')\n",
    "else: opts.use_gpu = 0;    print('gpu impossible!!')\n",
    "    \n",
    "fig = plt.figure(figsize=(15, 3))\n",
    "plt.subplot(1, 2, 1)\n",
    "plt.plot(epoch_losses)\n",
    "plt.plot(parameters.best_idx, epoch_losses[parameters.best_idx], 'ro')\n",
    "plt.title('loss')\n",
    "\n",
    "plt.subplot(1, 2, 2)\n",
    "plt.plot(all_F)\n",
    "plt.plot([parameters.best_idx, parameters.best_idx, parameters.best_idx], all_F[parameters.best_idx], 'ro')\n",
    "plt.title('F score')\n",
    "plt.legend(['train', 'dev', 'test'])\n",
    "plt.ylim([90.5, 100])\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "assert os.path.isfile(opts.test)\n",
    "assert parameters.tag_scheme in ['iob', 'iobes']\n",
    "\n",
    "if not os.path.isfile(opts.eval_script):\n",
    "    raise Exception('CoNLL evaluation script not found at \"%s\"' % eval_script)\n",
    "if not os.path.exists(opts.eval_temp):\n",
    "    os.makedirs(opts.eval_temp)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "test_sentences = load_sentences(opts.test, lower, zeros)\n",
    "update_tag_scheme(test_sentences, tag_scheme)\n",
    "test_data = prepare_dataset(\n",
    "    test_sentences, word_to_id, char_to_id, tag_to_id, lower\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "BiLSTM_CRF(\n",
       "  (char_embeds): Embedding(85, 25)\n",
       "  (char_cnn3): Conv2d(1, 25, kernel_size=(3, 25), stride=(1, 1), padding=(2, 0))\n",
       "  (word_embeds): Embedding(400176, 100)\n",
       "  (dropout): Dropout(p=0.5)\n",
       "  (lstm): LSTM(125, 200, bidirectional=True)\n",
       "  (hw_trans): Linear(in_features=25, out_features=25, bias=True)\n",
       "  (hw_gate): Linear(in_features=25, out_features=25, bias=True)\n",
       "  (h2_h1): Linear(in_features=400, out_features=200, bias=True)\n",
       "  (tanh): Tanh()\n",
       "  (hidden2tag): Linear(in_features=400, out_features=19, bias=True)\n",
       ")"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model = torch.load(opts.model_name)\n",
    "model_name = opts.model_name.split('/')[-1].split('.')[0]\n",
    "\n",
    "if opts.use_gpu:\n",
    "    model.cuda()\n",
    "model.eval()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "def evaluating(model, datas, display_confusion_matrix = False):\n",
    "    start_time = datetime.now();print('training start -%s' % start_time);\n",
    "    prediction = []\n",
    "    save = False\n",
    "    new_F = 0.0\n",
    "    confusion_matrix = torch.zeros((len(tag_to_id) - 2, len(tag_to_id) - 2))  # tag들의 갯수 - 2(start, stop)\n",
    "    for data in datas:\n",
    "        ground_truth_id = data['tags']\n",
    "        words    = data['str_words'] # ['EU', 'rejects', 'German', 'call', 'to', 'boycott', 'British', 'lamb', '.']\n",
    "        chars2   = data['chars']     # [[36, 58], [7, 1, 62, 1, 12, 3, 8], ...\n",
    "        caps     = data['caps']      # [1, 0, 2, 0, 0, 0, 2, 0, 0]\n",
    "\n",
    "        ## char embedding\n",
    "        if opts.char_mode == 'LSTM':\n",
    "            chars2_sorted = sorted(chars2, key=lambda p: len(p), reverse=True)\n",
    "            matching_char = {}\n",
    "            for i, ci in enumerate(chars2):\n",
    "                for j, cj in enumerate(chars2_sorted):\n",
    "                    if ci == cj and not j in matching_char and not i in matching_char.values():\n",
    "                        matching_char[j] = i\n",
    "                        continue\n",
    "            chars2_length = [len(c) for c in chars2_sorted]\n",
    "            char_maxl = max(chars2_length)\n",
    "            chars2_mask = np.zeros((len(chars2_sorted), char_maxl), dtype='int')\n",
    "            for i, c in enumerate(chars2_sorted):\n",
    "                chars2_mask[i, :chars2_length[i]] = c\n",
    "            chars2_mask = Variable(torch.LongTensor(chars2_mask))\n",
    "\n",
    "        if opts.char_mode == 'CNN':\n",
    "            matching_char = {}\n",
    "            chars2_length = [len(c) for c in chars2]\n",
    "            char_maxl = max(chars2_length)\n",
    "            chars2_mask = np.zeros((len(chars2_length), char_maxl), dtype='int')\n",
    "            for i, c in enumerate(chars2):\n",
    "                chars2_mask[i, :chars2_length[i]] = c\n",
    "            chars2_mask = Variable(torch.LongTensor(chars2_mask))\n",
    "        \n",
    "        eval_sentence_in = Variable(torch.LongTensor(data['words']))\n",
    "        eval_caps  = Variable(torch.LongTensor(caps))\n",
    "        # 변수들 gpu로 보내기\n",
    "        if opts.use_gpu:\n",
    "            eval_sentence_in= eval_sentence_in.cuda()\n",
    "            chars2_mask     = chars2_mask.cuda()\n",
    "            eval_caps       = eval_caps.cuda()\n",
    "        ############################################################################\n",
    "        ## !! inference!!\n",
    "        ## input : 문장, caption, 단어들, 단어길이들, 단어매칭dict\n",
    "        ## output: score : tensor(237.0103, device='cuda:0', grad_fn=<SelectBackward>)\n",
    "        ##         target_sequence : [4, 0, 5, 0, 0, 0, 5, 0, 0]\n",
    "        ############################################################################\n",
    "        score, predict_target_sequence = model(sentence= eval_sentence_in,\n",
    "                                               caps    = eval_caps,                             \n",
    "                                               chars        = chars2_mask,\n",
    "                                               chars2_length= chars2_length,\n",
    "                                               matching_char= matching_char)  # target이 없다.\n",
    "        predicted_id = predict_target_sequence    \n",
    "        # word, true_id, pred_id\n",
    "        # 단어,    정답, 예측값\n",
    "        # EU       4     4\n",
    "        # rejects  0     0\n",
    "        # German   5     5\n",
    "        # call     0     0\n",
    "        # to       0     0\n",
    "        # boycott  0     0\n",
    "        # British  5     5\n",
    "        # lamb     0     0\n",
    "        # .        0     0\n",
    "        for (word, true_id, pred_id) in zip(words, ground_truth_id, predicted_id):\n",
    "            line = ' '.join([word, id_to_tag[true_id], id_to_tag[pred_id]])\n",
    "            prediction.append(line)\n",
    "            confusion_matrix[true_id, pred_id] += 1 # 17 x 17 행렬에서, [실제값, 예측값] 부분에 +1 추가\n",
    "        prediction.append('')\n",
    "    predf  = opts.eval_temp + '/pred.' + opts.name\n",
    "    scoref = opts.eval_temp + '/score.' + opts.name \n",
    "    \n",
    "    ## inference 출력!!!! - /evaluation/temp/pred.test\n",
    "    ## word   true   prediction\n",
    "    ## prediction\n",
    "    # EU S-ORG O\n",
    "    # rejects O O\n",
    "    # German S-MISC O\n",
    "    # call O O\n",
    "    with open(predf, 'w') as f:\n",
    "        f.write('\\n'.join(prediction))\n",
    "    # evaluation 폴더에 있는 conlleval 파일을 실행시킨다.\n",
    "    # 공식으로 제공하는 평가 및 저장 코드가 있다.\n",
    "    # system 명령어로 구동 - 'evaluation/conlleval < evaluation/temp/pred.test > evaluation/temp/score.test'\n",
    "    os.system('%s < %s > %s' % (opts.eval_script, predf, scoref))\n",
    "    ############################################################################\n",
    "    ## display\n",
    "    ############################################################################\n",
    "    # 위 명령어로 평가가 진행되고 'scoref' 경로에 socre가 저장된다\n",
    "    # best f1 score를 갱신하면, save=True를 return 하고 저장하도록 한다\n",
    "    eval_lines = [l.rstrip() for l in codecs.open(scoref, 'r', 'utf8')]\n",
    "    for i, line in enumerate(eval_lines):\n",
    "        print(line)\n",
    "#         if i == 1:\n",
    "#             new_F = float(line.strip().split()[-1])\n",
    "#             if new_F > best_F:\n",
    "#                 best_F = new_F\n",
    "#                 save = True\n",
    "#                 print('the best F is ', new_F)\n",
    "    ############################################################################\n",
    "    ## display, confusion matrix\n",
    "    ############################################################################\n",
    "    if display_confusion_matrix:\n",
    "        print((\"{: >2}{: >7}{: >7}%s{: >9}\" % (\"{: >6}\" * confusion_matrix.size(0))).format(\n",
    "                \"ID\", \"NE\", \"Total\",\n",
    "                *([id_to_tag[i] for i in range(confusion_matrix.size(0))] + [\"Percent\"])\n",
    "                ))\n",
    "        for i in range(confusion_matrix.size(0)):\n",
    "            print((\"{: >2}{: >7}{: >7}%s{: >9}\" % (\"{: >6}\" * confusion_matrix.size(0))).format(\n",
    "                    str(i), id_to_tag[i], str(int(confusion_matrix[i].sum())),\n",
    "                    *([int(confusion_matrix[i][j]) for j in range(confusion_matrix.size(0))] +\n",
    "                      [\"%.2f\" % (confusion_matrix[i][i] * 100. / max(1, confusion_matrix[i].sum()))])\n",
    "                    ))\n",
    "    return start_time"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "training start -2019-06-05 14:11:46.029094\n",
      "processed 46435 tokens with 5628 phrases; found: 5665 phrases; correct: 5121.\n",
      "accuracy:  97.82%; precision:  90.40%; recall:  90.99%; FB1:  90.69\n",
      "              LOC: precision:  91.61%; recall:  93.92%; FB1:  92.75  1704\n",
      "             MISC: precision:  80.17%; recall:  81.67%; FB1:  80.91  706\n",
      "              ORG: precision:  88.42%; recall:  88.04%; FB1:  88.23  1649\n",
      "              PER: precision:  95.64%; recall:  94.99%; FB1:  95.31  1606\n",
      "ID     NE  Total     O S-LOC B-PER E-PER S-ORGS-MISC B-ORG E-ORG S-PER I-ORG B-LOC E-LOCB-MISCE-MISCI-MISC I-PER I-LOC  Percent\n",
      " 0      O  38323 38043    17     6     5    32    28    21    31    11    21     6     8    36    45    10     0     3    99.27\n",
      " 1  S-LOC   1436    11  1344     2     0    29    16     7     4     3     0     3     2    12     1     0     0     2    93.59\n",
      " 2  B-PER   1086     3     0  1075     0     0     0     5     0     0     0     1     0     1     0     0     1     0    98.99\n",
      " 3  E-PER   1086     1     0     0  1071     0     0     0     4     2     1     0     1     0     1     0     5     0    98.62\n",
      " 4  S-ORG   1082    42    22     0     0   967    17     9    10    10     0     1     0     3     0     1     0     0    89.37\n",
      " 5 S-MISC    525    19     8     1     0    24   445     5     0     7     0     3     0     6     7     0     0     0    84.76\n",
      " 6  B-ORG    579     7    11     7     0     1     2   498     0     2     9    32     0    10     0     0     0     0    86.01\n",
      " 7  E-ORG    579    19     1     0     5     1     0     0   501     2     9     0    29     0    11     0     1     0    86.53\n",
      " 8  S-PER    531    14    24     0     7    17     0     2     1   464     0     0     2     0     0     0     0     0    87.38\n",
      " 9  I-ORG    256     9     1     0     2     0     0     6     3     0   217     1     4     1     0     5     2     5    84.77\n",
      "10  B-LOC    232     0     1     2     0     0     0    10     0     0     0   217     0     1     0     0     0     1    93.53\n",
      "11  E-LOC    232     1     1     0     2     0     0     0     7     0     3     0   216     0     1     0     0     1    93.10\n",
      "12 B-MISC    177    12     0     5     0     3     5    10     0     4     1     9     0   123     1     4     0     0    69.49\n",
      "13 E-MISC    177    19     1     0     5     1     0     0    11     2     0     0     9     0   128     1     0     0    72.32\n",
      "14 I-MISC     39     3     0     0     0     0     0     1     1     0     2     0     0     3     1    28     0     0    71.79\n",
      "15  I-PER     70     0     0     0     0     0     0     0     0     0     2     0     0     0     0     1    67     0    95.71\n",
      "16  I-LOC     25     1     0     0     0     0     0     0     0     0     2     0     1     0     0     1     0    20    80.00\n",
      "\n",
      "소요시간 : 0:00:18.421192\n"
     ]
    }
   ],
   "source": [
    "start_time = evaluating(model, test_data, display_confusion_matrix=True)\n",
    "print('\\n소요시간 : %s' % ((datetime.now() - start_time)))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "hide_input": false,
  "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": "3.6.7"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {},
   "toc_section_display": true,
   "toc_window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}