wikitext_103.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "from fastai import *\n",
    "from fastai.text import *"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Wikitext 103"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "PATH = Path('../data/wikitext-103/60k')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "def read_file(filename):\n",
    "    with open(PATH/f'{filename}.txt', encoding='utf8') as f:\n",
    "        text = f.readlines()\n",
    "    df = pd.DataFrame({'text':np.array(text), 'labels':np.zeros(len(text))}, columns=['labels', 'text'])\n",
    "    df.to_csv(PATH/f'{filename}.csv', header=False, index=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "read_file('valid')\n",
    "#read_file('test')\n",
    "read_file('train')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tokenizing train. This might take a while so you should grab a coffee.\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, max=181), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Numericalizing train.\n",
      "Tokenizing valid. This might take a while so you should grab a coffee.\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, max=1), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Numericalizing valid.\n"
     ]
    }
   ],
   "source": [
    "tokenizer = Tokenizer(rules=rules, special_cases=[BOS, FLD, 'xxunk', 'xxpad'])\n",
    "train_ds, valid_ds = TextDataset.from_csv(PATH, tokenizer, max_vocab=60000)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "bs,bptt = 80,70\n",
    "train_dl = LanguageModelLoader(np.concatenate(train_ds.ids), bs, bptt)\n",
    "valid_dl = LanguageModelLoader(np.concatenate(valid_ds.ids), bs, bptt)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "vocab_size = len(train_ds.vocab.itos)\n",
    "data = DataBunch(train_dl, valid_dl)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "emb_sz, nh, nl = 400, 1150, 3\n",
    "dps = np.array([0.25, 0.1, 0.2, 0.02, 0.15]) * 0.1\n",
    "model = get_language_model(vocab_size, emb_sz, nh, nl, 0, input_p=dps[0], output_p=dps[1], weight_p=dps[2], \n",
    "                           embed_p=dps[3], hidden_p=dps[4])\n",
    "learn = Learner(data, model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "learn.opt_fn = partial(optim.Adam, betas=(0.8,0.99))\n",
    "learn.callbacks.append(RNNTrainer(learn, bptt, alpha=2, beta=1))\n",
    "learn.callback_fns.append(partial(GradientClipping, clip=0.12))\n",
    "learn.metrics = [accuracy]\n",
    "learn.true_wd=False"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "VBox(children=(HBox(children=(IntProgress(value=0, max=10), HTML(value=''))), HTML(value='epoch  train loss  v…"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Total time: 10:28:58\n",
      "epoch  train loss  valid loss  accuracy\n",
      "0      3.867067    3.917060    0.357719  (1:03:08)\n",
      "1      3.736901    3.824207    0.363141  (1:03:02)\n",
      "2      3.734958    3.847070    0.360801  (1:02:49)\n",
      "3      3.690017    3.810396    0.364572  (1:02:50)\n",
      "4      3.615047    3.704101    0.374547  (1:02:58)\n",
      "5      3.555377    3.589375    0.387541  (1:02:58)\n",
      "6      3.526703    3.495668    0.399273  (1:02:42)\n",
      "7      3.531244    3.480701    0.401607  (1:02:47)\n",
      "8      3.505409    3.472463    0.402647  (1:02:46)\n",
      "9      3.501862    3.468110    0.403324  (1:02:53)\n",
      "\n"
     ]
    }
   ],
   "source": [
    "fit_one_cycle(learn, 10, 5e-3, (0.8,0.7), wd=1e-7)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "learn.save('lstm_wt103')"
   ]
  }
 ],
 "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": "3.7.0"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": false,
   "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
}