Skip to content
Snippets Groups Projects
01-Distilbert.ipynb 15.1 KiB
Newer Older
Jean-Luc Parouty's avatar
Jean-Luc Parouty committed
{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<img width=\"800px\" src=\"../fidle/img/header.svg\"></img>\n",
Jean-Luc Parouty's avatar
Jean-Luc Parouty committed
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509
    "\n",
    "# <!-- TITLE --> [TRANS1] - IMDB, Sentiment analysis with Transformers \n",
    "<!-- DESC --> Using a Tranformer to perform a sentiment analysis (IMDB) - Jean Zay version\n",
    "<!-- AUTHOR : Hatim Bourfoune (IDRIS) and Nathan Cassereau (IDRIS) -->\n",
    "\n",
    "By : Hatim Bourfoune (IDRIS) and Nathan Cassereau (IDRIS)\n",
    "\n",
    "\n",
    "## Objectives :\n",
    " - Complement the learning of a Transformer to perform a sentiment analysis\n",
    " - Understand the use of a pre-trained transformer\n",
    "\n",
    "This task is exactly the same as the Sentiment analysis with text embedding. Only this time, \n",
    "we are going to exploit the strenght of transformers. Considering how computation-heavy transformer \n",
    "pretraining is, we are going to use a pretrained BERT model from HuggingFace. \n",
    "This notebook performs the fine-tuning process. If possible, try to use a GPU to speed up \n",
    "the training, transformers are difficult to train on CPU.\n",
    "\n",
    "## What we are going to do:\n",
    "\n",
    "* Retrieve the dataset\n",
    "* Prepare the dataset\n",
    "* Fetch a pretrained BERT model from HuggingFace's platform (https://huggingface.co/models)\n",
    "* Fine-tune the model on a sequence classification task: the sentiment analysis of the IMDB dataset\n",
    "* Evaluate the result\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Installations\n",
    "\n",
    "**IMPORTANT :** We will need to use the library `transformers` created by HuggingFace.\n",
    "\n",
    "The next line only applies on Jean Zay, it allows us to load a very specific environment, which contains Tensorflow with GPU support. Ignore that line if this notebook is not executed on Jean Zay."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "50QKMUQzPv3n",
    "outputId": "3ac2016d-596d-4f9a-c2ec-738c939c49a0"
   },
   "outputs": [],
   "source": [
    "#!pip install transformers\n",
    "!module load tensorflow-gpu/py3/2.6.0"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Imports and initialisation "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "ZrV8ONYZPi8L",
    "outputId": "ad10d385-3e1f-4ecf-80f2-87dccc286db7"
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "\n",
    "import tensorflow as tf\n",
    "import tensorflow.keras as keras\n",
    "import tensorflow.keras.datasets.imdb as imdb\n",
    "from tensorflow.keras.layers import Dense, Dropout\n",
    "from tensorflow.keras.optimizers import Adam\n",
    "from tensorflow.keras.losses import SparseCategoricalCrossentropy\n",
    "from tensorflow.keras.metrics import SparseCategoricalAccuracy\n",
    "from tensorflow.keras import mixed_precision\n",
    "\n",
    "from transformers import (\n",
    "    DistilBertTokenizer,\n",
    "    TFDistilBertModel,\n",
    "    DataCollatorWithPadding,\n",
    "    BertTokenizer,\n",
    "    TFBertModel\n",
    ")\n",
    "\n",
    "import pickle\n",
    "import multiprocessing\n",
    "import itertools\n",
    "import os\n",
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "\n",
    "print(\"Tensorflow \", tf.__version__)\n",
    "n_gpus = len(tf.config.list_physical_devices('GPU'))\n",
    "print(\"#GPUs: \", n_gpus)\n",
    "if n_gpus > 0:\n",
    "    !nvidia-smi -L\n",
    "policy = mixed_precision.Policy('mixed_float16')\n",
    "mixed_precision.set_global_policy(policy)\n",
    "os.environ[\"TOKENIZERS_PARALLELISM\"] = \"true\"\n",
    "\n",
    "np.random.seed(987654321)\n",
    "tf.random.set_seed(987654321)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Parameters\n",
    "\n",
    "* `vocab_size` refers to the number of words which will be remembered in our vocabulary.\n",
    "* `hide_most_frequently` is the number of ignored words, among the most common ones.\n",
    "* `review_len` is the review length.\n",
    "* `n_cpus` is the number of CPU which will be used for data preprocessing.\n",
    "* `distil` refers to whether or not we are going to use a DistilBert model or a regular Bert model.\n",
    "* `load_locally` will fetch data locally, otherwise will download on the Internet (requires an Internet connection, not possible on Jean Zay)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "FhIuZkS2PnTE"
   },
   "outputs": [],
   "source": [
    "vocab_size = 30000\n",
    "hide_most_frequently = 0\n",
    "\n",
    "review_len = 512\n",
    "\n",
    "epochs = 1\n",
    "batch_size = 32\n",
    "\n",
    "fit_verbosity = 1\n",
    "scale = 1\n",
    "\n",
    "n_cpus = 6\n",
    "distil = True\n",
    "load_locally = True # if set to False, will fetch data from the internet (requires an internet connection)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Retrieve the dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "qaRtDy9wQinS",
    "outputId": "9d2d9e12-74fb-4eee-9d9b-d4a7148e2dc2"
   },
   "outputs": [],
   "source": [
    "if load_locally:\n",
    "    with open(\"dataset\", \"rb\") as file_:\n",
    "        (x_train, y_train), (x_test, y_test) = pickle.load(file_)\n",
    "else:\n",
    "    (x_train, y_train), (x_test, y_test) = imdb.load_data(\n",
    "        num_words=vocab_size,\n",
    "        skip_top=hide_most_frequently,\n",
    "        seed=123456789,\n",
    "    )\n",
    "    with open(\"dataset\", \"wb\") as file_:\n",
    "        pickle.dump(((x_train, y_train), (x_test, y_test)), file_)\n",
    "\n",
    "\n",
    "y_train = np.asarray(y_train).astype('float32')\n",
    "y_test  = np.asarray(y_test ).astype('float32')\n",
    "\n",
    "n1 = int(scale * len(x_train))\n",
    "n2 = int(scale * len(x_test))\n",
    "x_train, y_train = x_train[:n1], y_train[:n1]\n",
    "x_test,  y_test  = x_test[:n2],  y_test[:n2]\n",
    "\n",
    "print(\"x_train : {}  y_train : {}\".format(x_train.shape, y_train.shape))\n",
    "print(\"x_test  : {}  y_test  : {}\".format(x_test.shape,  y_test.shape))\n",
    "print('\\nReview sample (x_train[12]) :\\n\\n',x_train[12])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "nbF1uktpRdXy"
   },
   "outputs": [],
   "source": [
    "if load_locally:\n",
    "    with open(\"word_index\", \"rb\") as file_:\n",
    "        word_index = pickle.load(file_)\n",
    "else:\n",
    "    word_index = imdb.get_word_index()\n",
    "    with open(\"word_index\", \"wb\") as file_:\n",
    "        pickle.dump(word_index, file_)\n",
    "\n",
    "word_index = {w:(i+3) for w,i in word_index.items()}\n",
    "word_index.update({'[PAD]':0, '[CLS]':1, '[UNK]':2})\n",
    "index_word = {index:word for word,index in word_index.items()} \n",
    "\n",
    "# Add a nice function to transpose:\n",
    "def dataset2text(review):\n",
    "    return ' '.join([index_word.get(i, \"?\") for i in review[1:]])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "DifmZNsKR38n",
    "outputId": "cb5f9819-1930-478d-f3f2-45f06e04c5d4"
   },
   "outputs": [],
   "source": [
    "print(dataset2text(x_train[12]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Fetch the model from HuggingFace"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def load_model(distil, load_locally):\n",
    "    if load_locally:\n",
    "        if distil:\n",
    "            bert_model = TFDistilBertModel.from_pretrained(\"distilbert_model\")\n",
    "            tokenizer = DistilBertTokenizer(\"distilbert_vocab.txt\", do_lower_case=True)\n",
    "        else:\n",
    "            bert_model = TFBertModel.from_pretrained(\"bert_model\")\n",
    "            tokenizer = BertTokenizer(\"bert_vocab.txt\", do_lower_case=True)\n",
    "        return bert_model, tokenizer\n",
    "\n",
    "    if distil:\n",
    "        bert_model = TFDistilBertModel.from_pretrained(\"distilbert-base-uncased\")\n",
    "        tokenizer = DistilBertTokenizer.from_pretrained(\"distilbert-base-uncased\")\n",
    "        bert_model.save_pretrained(\"distilbert_model\")\n",
    "        tokenizer.save_vocabulary(\"distilbert_vocab.txt\")\n",
    "    else:\n",
    "        bert_model = TFBertModel.from_pretrained(\"bert-base-uncased\")\n",
    "        tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n",
    "        bert_model.save_pretrained(\"bert_model\")\n",
    "        tokenizer.save_vocabulary(\"bert_vocab.txt\")\n",
    "    return bert_model, tokenizer\n",
    "\n",
    "bert_model, tokenizer = load_model(distil, load_locally)\n",
    "bert_model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Prepare the dataset "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "KKwI-RIXnWWd",
    "outputId": "42dc5943-b060-4a08-9180-cd5914980527"
   },
   "outputs": [],
   "source": [
    "def tokenize_sample(sample, tokenizer):\n",
    "    return tokenizer(dataset2text(sample), truncation=True, max_length=review_len)\n",
    "\n",
    "def distributed_tokenize_dataset(dataset):\n",
    "    ds = list(dataset)\n",
    "    with multiprocessing.Pool(n_cpus) as pool:\n",
    "        tokenized_ds = pool.starmap(\n",
    "            tokenize_sample,\n",
    "            zip(ds, itertools.repeat(tokenizer, len(ds)))\n",
    "        )\n",
    "    return tokenized_ds\n",
    "\n",
    "tokenized_x_train = distributed_tokenize_dataset(x_train)\n",
    "tokenized_x_test = distributed_tokenize_dataset(x_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "TivZYh8vnZlS"
   },
   "outputs": [],
   "source": [
    "data_collator = DataCollatorWithPadding(tokenizer, return_tensors=\"tf\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "7Up0APYtwFm7",
    "outputId": "37cb98bd-a0d3-47c2-9f91-96f94abf4b2e"
   },
   "outputs": [],
   "source": [
    "data_collator(tokenized_x_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "U6Lhjfh6maIF",
    "outputId": "f4798e7f-bc69-47fe-e2a7-c0155a99cca7"
   },
   "outputs": [],
   "source": [
    "def make_dataset(x, y):\n",
    "    collated = data_collator(x)\n",
    "    dataset = tf.data.Dataset.from_tensor_slices(\n",
    "        (collated['input_ids'], collated['attention_mask'], y)\n",
    "    )\n",
    "    transformed_dataset = (\n",
    "        dataset\n",
    "        .map(\n",
    "            lambda x, y, z: ((x, y), z)\n",
    "        )\n",
    "        .shuffle(25000)\n",
    "        .batch(batch_size)\n",
    "    )\n",
    "    return transformed_dataset\n",
    "\n",
    "train_ds = make_dataset(tokenized_x_train, y_train)\n",
    "test_ds = make_dataset(tokenized_x_test, y_test)\n",
    "\n",
    "for x, y in train_ds:\n",
    "    print(x)\n",
    "    break"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Add a new head to the model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "class ClassificationModel(keras.Model):\n",
    "\n",
    "    def __init__(self, bert_model):\n",
    "        super(ClassificationModel, self).__init__()\n",
    "        self.bert_model = bert_model\n",
    "        self.pre_classifier = Dense(768, activation='relu')\n",
    "        self.dropout = Dropout(0.1)\n",
    "        self.classifier = Dense(2)\n",
    "\n",
    "    def call(self, x):\n",
    "        x = self.bert_model(x)\n",
    "        x = x.last_hidden_state\n",
    "        x = x[:, 0] # get the output of the classification token\n",
    "        x = self.pre_classifier(x)\n",
    "        x = self.dropout(x)\n",
    "        x = self.classifier(x)\n",
    "        return x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = ClassificationModel(bert_model)\n",
    "x = next(iter(train_ds))[0]\n",
    "model(x)\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Train! "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "4jDAzxxwXLT1",
    "outputId": "bc4d5f62-9fa7-426d-a9e2-2fa4d2bdf780"
   },
   "outputs": [],
   "source": [
    "model.compile(\n",
    "    optimizer=Adam(1e-05),\n",
    "    loss=SparseCategoricalCrossentropy(from_logits=True),\n",
    "    metrics=[SparseCategoricalAccuracy('accuracy')]\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 419
    },
    "id": "KtmfFjL02Ano",
    "outputId": "ca174c57-b8f9-4d50-a53a-03761556e492"
   },
   "outputs": [],
   "source": [
    "history = model.fit(\n",
    "    train_ds,\n",
    "    epochs=epochs,\n",
    "    verbose=fit_verbosity\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Evaluation "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "_, score = model.evaluate(test_ds)\n",
    "colors = sns.color_palette('pastel')[2:]\n",
    "accuracy_score = [score, 1 - score]\n",
    "plt.pie(\n",
    "    accuracy_score,\n",
    "    labels=[\"Accurate\", \"Mistaken\"],\n",
    "    colors=colors,\n",
    "    autopct=lambda val: f\"{val:.2f}%\",\n",
    "    explode=(0.0, 0.1)\n",
    ")\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "collapsed_sections": [],
   "name": "Untitled0.ipynb",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3.9.2 ('fidle-env')",
Jean-Luc Parouty's avatar
Jean-Luc Parouty committed
   "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": "3.9.2"
  },
  "vscode": {
   "interpreter": {
    "hash": "b3929042cc22c1274d74e3e946c52b845b57cb6d84f2d591ffe0519b38e4896d"
   }
Jean-Luc Parouty's avatar
Jean-Luc Parouty committed
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}