TP4_Transformers_et_affinage.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "ci1-MA27MWp5"
   },
   "source": [
    "# TP4 : Architecture Transformer et affinage\n",
    "\n",
    "<i>TP adapté de celui de Yongxin Zhou</i>\n",
    "\n",
    "**Date limite de rendu :** À déterminer, date provisoire : le 15 octobre 2024 à 10:30.\n",
    "\n",
    "Références :\n",
    "\n",
    "- Tutoriels dont ce TP est inspiré : [BertViz](https://colab.research.google.com/drive/1hXIQ77A4TYS4y3UthWF-Ci7V7vVUoxmQ?usp=sharing#scrollTo=T3H0qUZvPOP4), [HuggingFace](https://huggingface.co/docs/transformers/tasks/translation)\n",
    "- [Tutoriel Pytorch](https://pytorch.org/tutorials/intermediate/seq2seq_translation_tutorial.html)\n",
    "- [Cours général](https://github.com/nyu-dl/NLP_DL_Lecture_Note/blob/master/lecture_note.pdf)\n",
    "- [Documentation de HuggingFace](https://huggingface.co/docs/transformers/index)\n",
    "\n",
    "## Introduction\n",
    "\n",
    "Le but de ce TP est de donner un aperçu de l'utilisation de modèles pré-entraînés utilisant l'architecture Transformer, ici pour une tâche de traduction de texte.\n",
    "\n",
    "L'architecture Transformer, dont l'invention remonte à 2017 dans l'article [Attention Is All You Need](https://arxiv.org/abs/1706.03762), est depuis plusieurs années l'état de l'art de la traduction automatique neuronale, ainsi que de beaucoup d'autres tâches de traitement automatique des langues.\n",
    "\n",
    "Les modèles que nous allons étudier dans ce TP sont de type encoder-décodeur. Dans les Transformer, l'encodeur utilise un mécanisme dit d'auto-attention (*self-attention*) sur les couches d'entrée, tandis que le décodeur combine l'auto-attention avec l'attention sur l'encodeur. Contrairement aux RNN, aucune récurrence n'est utilisée, et le modèle peut digérer un texte entier en une seule itération. De ce fait, un modèle transformer est plus rapide à entraîner et à utiliser.\n",
    "\n",
    "## Préparatifs\n",
    "\n",
    "Comme pendant le TP précédent, n'oubliez pas de redémarrer le noyau Python après l'exécution de cette cellule, et de bien sélectionner une instance Colab avec un GPU. Pour ce TP, aucun fichier auxilliaire n'est nécessaire."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "FGZIjJCHskAe"
   },
   "outputs": [],
   "source": [
    "# Installation des bibliothèques. Bertviz est un outil pour la visualisation\n",
    "# de l'attention dans les modèles de type Transformer (BERT, GPT-2, Albert, XLNet, RoBERTa, CTRL, etc.)\n",
    "!pip install transformers[torch] datasets evaluate sacrebleu bertviz"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "Pthp157SfmZn"
   },
   "source": [
    "Importons désormais les bibliothèques principales dont nous aurons besoin par la suite."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "OJTZpCGAgPqy"
   },
   "outputs": [],
   "source": [
    "# Datasets : Gestion des jeux de données, pour une utilisation facile avec la bibliothèque transformers\n",
    "# Transformers : Gestion de modèles neuronaux d'architecture Transformer\n",
    "# Bertviz : Visualisation de l'attention dans un modèle Transformer\n",
    "# Evaluate : Évaluation des performances du modèle\n",
    "# Matplotlib : Visualisation sous forme de graphiques\n",
    "# Numpy : Gestion des matrices et des données numériques\n",
    "\n",
    "from datasets import load_dataset\n",
    "from transformers import T5Tokenizer, AutoTokenizer\n",
    "from transformers import DataCollatorForSeq2Seq\n",
    "from transformers import AutoModelForSeq2SeqLM, Seq2SeqTrainingArguments, Seq2SeqTrainer\n",
    "from transformers import pipeline\n",
    "from transformers import T5ForConditionalGeneration, T5Tokenizer\n",
    "from bertviz import head_view, model_view\n",
    "\n",
    "import evaluate\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "import numpy as np\n",
    "import random"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "BlFaJTHxLrYx"
   },
   "source": [
    "# Corpus de traduction et préparation\n",
    "\n",
    "Comme précédemment, nous allons travailler sur un système de traduction de l'anglais vers le français. Pour cela, on utilisera le jeu de données [OPUS book](https://huggingface.co/datasets/opus_books). Il s'agit d'une collection de textes alignés en plusieurs langues. On rappelle que pour entraîner un modèle de traduction, il est nécessaire d'avoir des données parallèles, c'est-à-dire d'une collection de phrases sources et de traductions de référence alignées phrase par phrase et stockées dans deux fichiers, de telle sorte que chaque ligne du fichier de référence soit la traduction de la même ligne dans le fichier source.\n",
    "Dans cet exemple, on télécharge et prétraite les données parallèles à partir du [hub de HuggingFace](https://huggingface.co/). Il s'agit d'une plateforme en ligne stockant de nombreux jeux de données et de modèles Transformer pré-entrainés. La bibliothèque `transformers` gère le chargement des modèles, tandis que la bibiliothèque `datasets` perlet de charger les jeux de données.\n",
    "\n",
    "Ici, pour pouvoir travailler plus rapidement, on réduit la taille du jeu de données.\n",
    "On pourra par la suite changer la valeur dans le code ci-dessous pour avoir un jeu de données plus grand.\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "EIaUPEt2g0LZ"
   },
   "outputs": [],
   "source": [
    "# Téléchargement du jeu de données OPUS books (sous-jeu de données anglais-français) depuis le hub\n",
    "books = load_dataset(\"opus_books\", \"en-fr\")\n",
    "\n",
    "# Sélection de maximum 3000 points de données, sélectionnés aléatoirement\n",
    "n_datapoints = len(books[\"train\"])\n",
    "n_datapoints_to_keep = min(n_datapoints, 3000)\n",
    "\n",
    "random_indices = random.sample(range(n_datapoints), n_datapoints_to_keep)\n",
    "dataset = books.filter(lambda example, idx: idx in random_indices, with_indices=True)\n",
    "\n",
    "# Séparation automatique en jeu d'entraînement et jeu de validation\n",
    "# Attention : Il ne s'agit pas de la fonction habituelle de `scikit-learn`,\n",
    "# mais de celle de `datasets`. Celle-ci modifie le jeu de données en place.\n",
    "dataset = dataset[\"train\"].train_test_split(test_size=0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "ehLgJ4jisbgK"
   },
   "outputs": [],
   "source": [
    "# Le jeu de données a été automatiquement divisé\n",
    "dataset"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "0SNHqEeFhBa1"
   },
   "source": [
    "Comme d'habitude, on peut inspecter manuellement certains examples de notre jeu de données. On peut bien voir grâce à leur champ `id` que ceux-ci sont sélectionnés aléatoirement."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "Hnsnw4N9hABL"
   },
   "outputs": [],
   "source": [
    "for i in range(4):\n",
    "  print(dataset[\"train\"][i])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "_tl_79NQoLJZ"
   },
   "source": [
    "# Modèle Transformer\n",
    "\n",
    "L'architecture [Transformer](https://arxiv.org/abs/1706.03762) est actuellement l'état de l'art pour beaucoup de tâches du TALN, et notamment pour la traduction automatique. Au sein de cette architecture, l'encodeur utilise en générale un bloc d'auto-attention sur les couches précédentes, tandis que le décodeur combine un bloc d'auto-attention ainsi qu'un bloc dit d'attention encodeur-décodeur.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "HA27nd0ToPlr"
   },
   "source": [
    "![Transformer_decoder.png](data:image/png;base64,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)\n",
    "\n",
    "*Image tirée de [The Illustrated Transformer](https://jalammar.github.io/illustrated-transformer/)*"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "HJUELHh5oRKy"
   },
   "source": [
    "La figure ci-dessus montre un bloc de l'encodeur et un bloc du décodeur. En pratique, l'encodeur comme le décodeur sont constitués de multiple blocs, \"empilés\" les uns sur les autres.\n",
    "\n",
    "On rappelle que l'architecture Transformer est principalement basée sur le mécanisme d'auto-attention, qui permet de mesurer la force de la relation entre chaque paire de tokens. En général, on n'a pas une seule, mais plusieurs têtes d'attention (*attention heads*) au sein de chaque bloc, chaque tête pouvant en théorie apprendre un type de relation différent. C'est ce qu'on appelle l'attention multi-tête (*multihead attention*).\n",
    "\n",
    "La force de la relation entre les différents tokens sur une attention multi-tête peut être facilement visualisée. Par exemple, voici une représentation de l'auto-attention pour la phrase \"The animal didn't cross the street because it was too tired\" dans un modèle de type Transformer :\n",
    "\n",
    "\n",
    "![Visualisation de l'attention](data:image/png;base64,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)\n",
    "\n",
    "## Questions\n",
    "Cette figure est extraite de [The Illustrated Transformer](https://jalammar.github.io/illustrated-transformer/). En vous aidant si besoin de cette page web, expliquez :\n",
    "1. Que représentent les séquences de tokens à gauche et à droite ?\n",
    "2. Que représentent les liens entre les tokens ?\n",
    "3. Que représente les differentes couleurs ?\n",
    "4. À quoi fait référence \"layer\" ?\n",
    "5. À votre avis, quelle tête gère ici la résolution de coréférences, c'est-à-dire le fait de savoir à quel nom correspond le pronom \"it\" ?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "U8rOA8AELIsO"
   },
   "source": [
    "**Réponses :**\n",
    "1. Les tokens de gauche représentent la phrase d'entrée. Ici, les tokens de droite représentent la même phrase car il s'agit de l'attention input-input.\n",
    "2. Les liens entre les tokens représentent la force des scores d'attention (leur somme sur toutes les têtes d'attention).\n",
    "3. Chaque couleur représente une tête d'attention différente.\n",
    "4. Ici, nous ne regardons que les têtes d'attention à la couche 5. Il est important de noter que chaque couche a ses propres têtes.\n",
    "5. Ici, il faut regarder les scores individuels des têtes (blocs de couleur à gauche). Le pronom \"it\" fait référence à \"animal\", mais avant de lire la suite de la phrase, on peut également penser qu'il fait référence à \"street\". Il faut donc trouver une tête donnant des scores forts pour ces deux mots - il s'agit de la tête rouge (tête n° 4)."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "T7wYyev9iV2-"
   },
   "source": [
    "# Modèle et tokenization\n",
    "\n",
    "Il existe aujourd'hui énormément de modèles utilisant l'architecture Transformer. L'image ci-dessous représente un arbre généalogique des différents sous-architectures. Elle est loin d'être exhaustive !\n",
    "\n",
    "![Arbre des modèles](https://github.com/Mooler0410/LLMsPracticalGuide/blob/main/imgs/tree.jpg?raw=true)\n",
    "\n",
    "<!--Nous allons ici travailler avec un modèle de langage pré-entrainé appelé FLAN-T5. Il s'agit d'une version dérivative de T5, une série de modèles open-source seq2seq (encodeur-décodeur) de type Transformer publiés par Google en 2019. Ces modèles ont été entraînés sur une grande diversité de tâches et de données textuelles.\n",
    "\n",
    "Les modèles FLAN-T5 sont une version affinée des modèles T5, qui sont affinés (*fine-tuned*) en particulier sur des jeux de données contenant des paires d'instructions et de réponses (*instruction fine-tuning*), une méthode désormais très courante qui permet en général d'améliorer les performances d'un modèle sur des tâches directes. Ces modèles sont plus récents, sortis en 2022.\n",
    "\n",
    "Nous travaillerons avec le modèle `FLAN-T5-small`, qui est la version la plus petite de FLAN-T5. Il contient néanmoins 60 million de paramètres !-->\n",
    "\n",
    "Nous allons ici travailler avec un modèle de langage pré-entrainé appelé T5. Il s'agit d'une série de modèles open-source seq2seq (encodeur-décodeur) de type Transformer publiés par Google en 2019. Ces modèles ont été entraînés sur une grande diversité de tâches et de données textuelles.\n",
    "\n",
    "Nous travaillerons avec le modèle `T5-small`, qui est la version la plus petite de T5. Il contient néanmoins 60 million de paramètres !\n",
    "\n",
    "Un des nombreux avantages de travailler avec un modèle pré-entraîné est que le tokenizer pour celui-ci a également déjà été entraîné. Nous pouvons donc le charger et l'utiliser directement ! Ici, la première étape va consister à évaluer ses performances de base.\n",
    "\n",
    "Pour utiliser un modèle multi-tâche de type seq2seq comme T5, on insère un prefix à nos examples, pour indiquer au modèle pre-entrainé la tâche que l'on souhaite réaliser."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "aJgUu503iaHF"
   },
   "outputs": [],
   "source": [
    "# Nom du modèle pre-entrainé que l'on va utiliser pour le moment\n",
    "model_checkpoint = \"t5-small\"\n",
    "\n",
    "# Chargement du tokenizer aussi pre-entrainé specialement pour ce modèle.\n",
    "# La classe `AutoTokenizer` détecte automatiquement le type de tokenizer (ici, un BPE).\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, padding=True)\n",
    "\n",
    "source_lang = \"en\"\n",
    "target_lang = \"fr\"\n",
    "# Le prefix donne une instruction au modèle pour effectuer la tâche que l'on souhaite\n",
    "prefix = \"Translate from English to French: \"\n",
    "\n",
    "def preprocess_function(examples):\n",
    "    # Avant de tokenizer le texte, on lui ajoute le préfixe\n",
    "    model_inputs = []\n",
    "    for ex in examples:\n",
    "        inputs = prefix + ex['translation'][source_lang] + '</s>'\n",
    "        targets = ex['translation'][target_lang] + '</s>'\n",
    "        model_inputs.append(tokenizer(inputs, text_target=targets, max_length=64, truncation=True))\n",
    "    return model_inputs\n",
    "\n",
    "# Application du pré-traitement à l'ensemble du dataset\n",
    "tokenized_books_train = preprocess_function(dataset[\"train\"])\n",
    "tokenized_books_dev = preprocess_function(dataset[\"test\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "0FBscRb9X4xk"
   },
   "source": [
    "## Chargement du modèle T5 pré-entrainé\n",
    "\n",
    "On utilise les fonctions de HuggingFace pour charger le modèle, en utilisant simplement le nom du modèle (ici `t5-small)`. Tout comme `AutoTokenizer`, la classe `AutoModel` détecte automatiquement le sous-type d'architecture Transformer dans le modèle et l'instancie correctement."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "gxb1ZWjSolHz"
   },
   "outputs": [],
   "source": [
    "model = AutoModelForSeq2SeqLM.from_pretrained(model_checkpoint).to(\"cuda:0\")\n",
    "model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "8sfDGK2xEbg8"
   },
   "source": [
    "## Questions\n",
    "\n",
    "Inspectez le modèle avec le code précédent.\n",
    "\n",
    "1. Quels composants de l'architecture Transformer pouvez-vous identifier ?\n",
    "2. Combien de blocs l'encodeur contient-il ? Et le décodeur ?\n",
    "\n",
    "**Réponses :**\n",
    "1. On peut observer la couche d'embedding, l'encodeur et le décodeur, et les différents blocs au sein de ceux-ci, avec comme sous-composants les couches de type perceptron (LayerFF) et les modules d'auto-attention et d'attention croisée.\n",
    "2. On a 6 blocs pour l'encodeur, et 6 pour le décodeur (bloc 0, et blocs 1 à 5)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "TfK4Tes7o2s3"
   },
   "source": [
    "# Affinage du modèle\n",
    "\n",
    "Tel quel, le modèle que nous avons téléchargé est capable de faire de la traduction avec de bonnes performances (que nous mesurerons un peu plus loin). En effet, celui-ci a été entraîné sur de grandes quantités de données et \"comprend\" assez bien le langage naturel. Nous souhaitons désormais utiliser notre jeu de données pour l'affiner, c'est-à-dire le spécialiser dans la tâche de traduction qui nous intéresse.\n",
    "\n",
    "Comme le modèle est déjà assez performant, cet affinement n'aura pas besoin de durer très longtemps; il s'agit simplement de modifier légèrement ses poids afin d'augmenter ses performances.\n",
    "\n",
    "#### Construction des lots (batches).\n",
    "Les lots sont automatiquement mélangés (shuffled) avant chaque époque.\n",
    "Pour cela, on utilise un Collator, qui performe automatiquement le batching et padding."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "irrII_FLM84z"
   },
   "outputs": [],
   "source": [
    "data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=model_checkpoint, max_length=64, padding=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "u8HMBbLKOgVM"
   },
   "source": [
    "\n",
    "Pour évaluer les performances d'un modèle (durant ou après l'entraînement), il est utile d'include une métrique correspondant à la tâche sur laquelle on souhaite évaluer le modèle. La bibliothèque [`evaluate`](https://huggingface.co/docs/evaluate/index) permet de facilement charger un certain nombre de métriques standard d'évaluation.\n",
    "\n",
    "Pour la traduction, on va utiliser la métrique [SacreBLEU](https://huggingface.co/spaces/evaluate-metric/sacrebleu) mentionnée au dernier TP. La documentation d'`evaluate` contient plus d'information sur la manière dont ont peu charger et calculer une métrique."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "8GN8vUtGOctX"
   },
   "outputs": [],
   "source": [
    "metric = evaluate.load(\"sacrebleu\")\n",
    "\n",
    "# Les fonctions de post-traitement dépendent grandement du type de modèle utilisé.\n",
    "# Ces détails techniques ne sont pas à mémoriser et peuvent être trouvés dans la documentation de chaque type de modèle.\n",
    "# Par exemple, pour T5 : https://huggingface.co/docs/transformers/en/model_doc/t5\n",
    "def postprocess_text(preds, labels):\n",
    "    preds = [pred.strip() for pred in preds]\n",
    "    labels = [[label.strip()] for label in labels]\n",
    "\n",
    "    return preds, labels\n",
    "\n",
    "\n",
    "def compute_metrics(eval_preds):\n",
    "    preds, labels = eval_preds\n",
    "    if isinstance(preds, tuple):\n",
    "        preds = preds[0]\n",
    "    decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)\n",
    "\n",
    "    labels = np.where(labels != -100, labels, tokenizer.pad_token_id)\n",
    "    decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)\n",
    "\n",
    "    decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)\n",
    "\n",
    "    result = metric.compute(predictions=decoded_preds, references=decoded_labels)\n",
    "    result = {\"bleu\": result[\"score\"]}\n",
    "\n",
    "    result = {k: round(v, 4) for k, v in result.items()}\n",
    "    return result"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "_yUEdxXfOSfK"
   },
   "source": [
    "## Programmation de l'apprentissage\n",
    "\n",
    "Comme toujours, l'apprentissage est itératif et s'effectue sur un certain nombre d'époques. Pour vérifier si l'apprentissage évolue positivement, nous utiliserons la mesure BLEU vue précédemment, conjointement avec la fonction de cout (*loss*), pour évaluer le modèle à chaque pas d'itération.\n",
    "\n",
    "Le code ci-dessous utilise les classes de HuggingFace, notamment la classe `Seq2SeqTrainer` qui contient les paramètres décrivant l'entraînement ou affinage que l'on souhaite exécuter.\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "VZsQst5hOVr7"
   },
   "outputs": [],
   "source": [
    "training_args = Seq2SeqTrainingArguments(\n",
    "    output_dir=\"opus_books_en_fr_model\",\n",
    "    evaluation_strategy=\"epoch\",\n",
    "    logging_strategy=\"epoch\",\n",
    "    save_strategy=\"epoch\",\n",
    "    learning_rate=5e-5,\n",
    "    per_device_train_batch_size=32,\n",
    "    per_device_eval_batch_size=16,\n",
    "    weight_decay=0.01,\n",
    "    save_total_limit=3,\n",
    "    num_train_epochs=12,\n",
    "    generation_max_length=64,\n",
    "    metric_for_best_model=\"eval_loss\",\n",
    "    load_best_model_at_end=True,\n",
    "    predict_with_generate=True,\n",
    "    fp16=True,\n",
    "    push_to_hub=False,\n",
    ")\n",
    "\n",
    "trainer = Seq2SeqTrainer(\n",
    "    model=model,\n",
    "    args=training_args,\n",
    "    train_dataset=tokenized_books_train,\n",
    "    eval_dataset=tokenized_books_dev,\n",
    "    tokenizer=tokenizer,\n",
    "    data_collator=data_collator,\n",
    "    compute_metrics=compute_metrics,\n",
    ")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "lyYPOKKwyfO0"
   },
   "source": [
    "## Evaluation du modèle pre-entrainé\n",
    "\n",
    "Le modèle T5 est déjà capable d'effectuer de bonnes prédictions de traductions sans affinage. On peut mesurer ses performances, en mesurant par exemple le score BLEU sur l'ensemble de test :"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "xEnS_YlOyee_"
   },
   "outputs": [],
   "source": [
    "results = trainer.evaluate()\n",
    "results['eval_bleu']"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "tdOAQJuzoqFu"
   },
   "source": [
    "Ici, le score obtenu est malgré tout faible, car il s'agit du plus petit modèle de la série T5. La petite taille d'un modèle pénalise en effet ses capacités, surtout lorsque celui-ci doit apprendre à effectuer un grand nombre de tâches à la fois !\n",
    "\n",
    "## Exécution de l'affinage\n",
    "\n",
    "Astuce : vous pouvez régler `epoch` sur une petite valeur (par exemple, 2) dans la cellule plus haut, et relancer la cellule ci-dessous plusieurs fois pour continuer à entraîner votre modèle (`train()` ne réinitialise pas le modèle)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "ZhC3jcq3ou87"
   },
   "outputs": [],
   "source": [
    "trainer.train()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "q1GpyHUStbhw"
   },
   "source": [
    "On peut visualiser les courbes d'apprentissage and utilisant l'historique de la classe d'entraînement d'HuggingFace :"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "9jN7Gol5tYvC"
   },
   "outputs": [],
   "source": [
    "# Récupération de l'historique\n",
    "log_history = trainer.state.log_history\n",
    "\n",
    "eval_losses = [entry['eval_loss'] for entry in log_history if 'eval_loss' in entry]\n",
    "eval_epochs = [entry['epoch'] for entry in log_history if 'eval_loss' in entry]\n",
    "\n",
    "train_losses = [entry['loss'] for entry in log_history if 'loss' in entry]\n",
    "train_epochs = [entry['epoch'] for entry in log_history if 'loss' in entry]\n",
    "\n",
    "# Affichage\n",
    "plt.plot(train_epochs, train_losses, label=\"Train Loss\")\n",
    "plt.plot(eval_epochs, eval_losses, label=\"Validation Loss\")\n",
    "plt.xlabel(\"Epochs\")\n",
    "plt.ylabel(\"Loss\")\n",
    "plt.title(\"Training and Validation Loss Curves\")\n",
    "plt.legend()\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "I9nI8qeyzhaD"
   },
   "source": [
    "## Evaluation du modèle affiné"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "w60s6WBTyHW8"
   },
   "outputs": [],
   "source": [
    "results = trainer.evaluate()\n",
    "results['eval_bleu']"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "a_g4giWg5CvX"
   },
   "source": [
    "## Question\n",
    "\n",
    "1. Le fine-tuning apporte un gain significatif en performance?\n",
    "\n",
    "**Réponse :** Dépend de votre machine. En général, on observe un léger gain de quelques points en score BLEU."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "ZuFNuqRg2_bq"
   },
   "source": [
    "## Prédiction avec le modèle affiné\n",
    "\n",
    "Huggingface simplifie le processus de prédiction en offrant une classe appelée `Pipeline`, qui permet de faire des prédictions à partir d'un modèle pre-entrainé."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "RkL-YbgW1pEE"
   },
   "outputs": [],
   "source": [
    "text = \"Translate from English to French: My neighbor's black cat eats mice.\"\n",
    "\n",
    "translator = pipeline(\"translation_en_to_fr\", model=model, tokenizer=tokenizer, device=0)\n",
    "translator(text)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "RXkf-D3ApCT6"
   },
   "source": [
    "## Test interactif du modèle et visualisation des matrices d'attention\n",
    "\n",
    "Le code suivant enregistre le modèle, puis propose deux visualisations interactives permettant d'observer le contenu des têtes d'attention à chaque couche grâce aux fonctions fournies par `bertviz`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "pKtINngoIrA1"
   },
   "outputs": [],
   "source": [
    "trainer.save_model(\"opus_books_en_fr_model\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "CU7d2B6O-g2v"
   },
   "outputs": [],
   "source": [
    "def visualize_encoder_decoder(source_text, predicted_text, model=T5ForConditionalGeneration.from_pretrained(\"opus_books_en_fr_model\"), tokenizer=T5Tokenizer.from_pretrained(\"t5-small\")):\n",
    "  inputs = tokenizer(\n",
    "      source_text,\n",
    "      return_tensors=\"pt\",\n",
    "  )\n",
    "\n",
    "  with tokenizer.as_target_tokenizer():\n",
    "      decoder_input_ids = tokenizer(predicted_text, return_tensors=\"pt\", add_special_tokens=True).input_ids\n",
    "\n",
    "  # On fournit à nouveau les sorties au modèle afin de lire les scores d'attention\n",
    "  out = model(**inputs, decoder_input_ids=decoder_input_ids, output_attentions=True, return_dict=True)\n",
    "  encoder_attentions = out.encoder_attentions\n",
    "  cross_attentions = out.cross_attentions\n",
    "  decoder_attentions = out.decoder_attentions\n",
    "\n",
    "  encoder_text = tokenizer.convert_ids_to_tokens(inputs.input_ids[0])\n",
    "  decoder_text = tokenizer.convert_ids_to_tokens(decoder_input_ids[0])\n",
    "\n",
    "  model_view(\n",
    "        encoder_attention=encoder_attentions,\n",
    "        decoder_attention=decoder_attentions,\n",
    "        cross_attention=cross_attentions,\n",
    "        encoder_tokens= encoder_text,\n",
    "        decoder_tokens = decoder_text\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "-coAq8LUCsK0"
   },
   "outputs": [],
   "source": [
    "# Visualisation des têtes d'attention pour le modèle entier; cliquez sur une tête pour plus de détails\n",
    "\n",
    "source_text = \"Translate from English to French: My neighbor's black cat eats mice.\"\n",
    "predicted_text = translator(source_text)[0][\"translation_text\"]\n",
    "\n",
    "visualize_encoder_decoder(source_text, predicted_text)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "ocuDAu6eJe3x"
   },
   "outputs": [],
   "source": [
    "def visualize_encoder_decoder_heads(source_text, predicted_text, model=T5ForConditionalGeneration.from_pretrained(\"opus_books_en_fr_model\"), tokenizer=T5Tokenizer.from_pretrained(\"t5-small\")):\n",
    "  inputs = tokenizer(\n",
    "      source_text,\n",
    "      return_tensors=\"pt\",\n",
    "  )\n",
    "\n",
    "  with tokenizer.as_target_tokenizer():\n",
    "      decoder_input_ids = tokenizer(predicted_text, return_tensors=\"pt\", add_special_tokens=True).input_ids\n",
    "\n",
    "  out = model(**inputs, decoder_input_ids=decoder_input_ids, output_attentions=True, return_dict=True)\n",
    "  encoder_attentions = out.encoder_attentions\n",
    "  cross_attentions = out.cross_attentions\n",
    "  decoder_attentions = out.decoder_attentions\n",
    "\n",
    "  encoder_text = tokenizer.convert_ids_to_tokens(inputs.input_ids[0])\n",
    "  decoder_text = tokenizer.convert_ids_to_tokens(decoder_input_ids[0])\n",
    "\n",
    "  head_view(\n",
    "        encoder_attention=encoder_attentions,\n",
    "        decoder_attention=decoder_attentions,\n",
    "        cross_attention=cross_attentions,\n",
    "        encoder_tokens= encoder_text,\n",
    "        decoder_tokens = decoder_text\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "9qosHWrMOp5j"
   },
   "outputs": [],
   "source": [
    "# Visualisation tête par tête. Vous pouvez cliquer sur les couleurs individuelles\n",
    "# pour sélectionner/déselectionner des têtes, et double-cliquer pour n'afficher qu'une tête.\n",
    "\n",
    "source_text = \"Translate from English to French: My neighbor's black cat eats mice.\"\n",
    "predicted_text = translator(source_text)[0][\"translation_text\"]\n",
    "\n",
    "visualize_encoder_decoder_heads(source_text, predicted_text)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "fcRS9qHUOv4q"
   },
   "source": [
    "## Question\n",
    "\n",
    "1. Variez le texte source et inspectez les têtes d'attention, notamment dans la deuxième visualisation et en mode \"cross\" (attention encodeur-décodeur). Quelles informations pouvez-vous conclure sur le comportement de certaines têtes ?\n",
    "\n",
    "**Réponse :** Il s'agit d'une question ouverte et subjective. On peut voir qu'en général, il est difficile de comprendre ce que font les têtes des couches profondes (1 et +); seules les têtes de surface peuvent être identifiées. Certaines têtes pointent toujours sur le mot correspondant à la traduction du mot en cours. D'autres têtes semblent s'activer sur le symbole \":\"; on peut imaginer que comme T5 a été entraîné sur différentes tâches avec différents préfixes, ces têtes permettent de mettre le modèle dans un mode correspondant à la tâche en cours (cad que le modèle sait désormais ce qu'il doit faire et regarde soudainement toute la séquence d'entrée)."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "jAf2yG5eGEud"
   },
   "source": [
    "# Entraînement à partir de zéro\n",
    "\n",
    "Pour le moment, nous avons vu comment affiner un modèle pré-entraîné. Il est également possible de pré-entraîner soi-même un modèle à partir de zéro. Ici, nous allons entraîner un modèle qui utilise la même architecture que l'encoder-decoder T5, mais qui n'est pas entraîné (les poids sont initialisés aléatoirement)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "gBnMWipQIGj6"
   },
   "outputs": [],
   "source": [
    "configuration = model.config\n",
    "\n",
    "# Création d'un nouveau modèle (poids aléatoires) en gardant la même configuration que le modèle précédent\n",
    "model_from_scratch = T5ForConditionalGeneration(configuration)\n",
    "\n",
    "# Nouvelle instance d'entraînement, mais qui réutilise le même jeu de données et paramètres d'entraînement\n",
    "trainer_from_scratch = Seq2SeqTrainer(\n",
    "    model=model_from_scratch,\n",
    "    args=training_args,\n",
    "    train_dataset=tokenized_books_train,\n",
    "    eval_dataset=tokenized_books_dev,\n",
    "    tokenizer=tokenizer,\n",
    "    data_collator=data_collator,\n",
    "    compute_metrics=compute_metrics,\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "nBG65XlYJTAC"
   },
   "source": [
    "La performance d'un modèle initialisé aléatoirement est, comme on peut l'imaginer, très mauvaise :"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "Q3hKWVNwIxDv"
   },
   "outputs": [],
   "source": [
    "results = trainer_from_scratch.evaluate()\n",
    "results['eval_bleu']"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "p-lmGYXEIy-a"
   },
   "source": [
    "Nous allons maintenant entraîner le modèle, puis observer à nouveau le résultat."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "RnxO6r3zJem5"
   },
   "outputs": [],
   "source": [
    "trainer_from_scratch.train()\n",
    "results = trainer_from_scratch.evaluate()\n",
    "results['eval_bleu']"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "j2IoYYseYu9C"
   },
   "source": [
    "## Question\n",
    "\n",
    "1. Que peut-on en conclure par rapport aux avantages et inconvénients de l'entraînement à partir de zéro, comparé au paradigme pré-entraînement + affinage ?\n",
    "\n",
    "**Réponse : L'entraînement à partir de zéro permet d'avoir un modèle avec des connaissances globales sans a priori sur la tâche à effectuer. Il est utile afin d'avoir des modèles de base pouvant ensuite être affinés sur différentes tâches. En revanche, l'affinage est bien moins coûteux et permet en général d'obtenir de meilleures performances."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "mn3kDqIjJ9uk"
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "provenance": []
  },
  "gpuClass": "standard",
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.10.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}