{ "cells": [ { "cell_type": "markdown", "id": "976841dc", "metadata": {}, "source": [ "## Preparación de un dataset\n", "\n", "Descargamos el dataset y lo preparamos para el entrenamiento. En el caso de ejemplo, usaremos toxic-teenage-relationships, que son frases que describen si un comporamiento es tóxico o sano. Tienen una campo de texto y un campo de etiqueta, que vale 1 si es tóxico y 0 si no lo es. Acumula 267 ejemplos de entrenamiento y 66 para testear." ] }, { "cell_type": "code", "execution_count": 1, "id": "b9a1f255", "metadata": {}, "outputs": [ { "data": { "text/plain": [ "{'label': 1,\n", " 'text': 'Mi amiga no puede subir videos a tik tok porque su pareja no le deja'}" ] }, "execution_count": 1, "metadata": {}, "output_type": "execute_result" } ], "source": [ "from datasets import load_dataset\n", "data_files = {\"train\": \"train.csv\", \"test\": \"test.csv\"}\n", "dataset = load_dataset(\"toxic-teenage-relationships\", data_files=data_files, sep=\";\")\n", "dataset['train'][100]" ] }, { "cell_type": "markdown", "id": "6d0c740a", "metadata": {}, "source": [ "Una vez cargado el dataset, se crea un tokenizador para procesar el texto e incluir una estrategia para el padding y el truncamiento. Para poder procesar el dataset en un solo paso, se utiliza el método dataset.map para preprocesar todo el dataset." ] }, { "cell_type": "code", "execution_count": 2, "id": "01673605", "metadata": {}, "outputs": [], "source": [ "from transformers import AutoTokenizer\n", "\n", "tokenizer = AutoTokenizer.from_pretrained(\"dccuchile/bert-base-spanish-wwm-cased\")\n", "\n", "\n", "def tokenize_function(examples):\n", " return tokenizer(examples[\"text\"], padding=\"max_length\", truncation=True)\n", "\n", "\n", "tokenized_datasets = dataset.map(tokenize_function, batched=True)" ] }, { "cell_type": "markdown", "id": "08aacc14", "metadata": {}, "source": [ "Ahora vamos a convertir el dataset en formator de TensorFlow. Para eso usamos DefaultDataCollator, que junta los tensores en un batch para que el modelo se entrene en él. Debemos especificar el argumento return_tensors=\"tf\". \n" ] }, { "cell_type": "code", "execution_count": 3, "id": "4a854ead", "metadata": {}, "outputs": [], "source": [ "from transformers import DefaultDataCollator\n", "data_collator = DefaultDataCollator(return_tensors=\"tf\")" ] }, { "cell_type": "markdown", "id": "06346bc5", "metadata": {}, "source": [ "guardamos los dataset de train y de test\n" ] }, { "cell_type": "code", "execution_count": 4, "id": "698a98ca", "metadata": {}, "outputs": [], "source": [ "train_dataset = tokenized_datasets[\"train\"]\n", "eval_dataset = tokenized_datasets[\"test\"]" ] }, { "cell_type": "markdown", "id": "38a6c521", "metadata": {}, "source": [ "\n", "\n", "En primer lugar, vamos a crear el modelo\n", "\n" ] }, { "cell_type": "code", "execution_count": 5, "id": "843f218d", "metadata": { "scrolled": true }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "All model checkpoint layers were used when initializing TFBertForSequenceClassification.\n", "\n", "Some layers of TFBertForSequenceClassification were not initialized from the model checkpoint at dccuchile/bert-base-spanish-wwm-cased and are newly initialized: ['classifier', 'bert/pooler/dense/kernel:0', 'bert/pooler/dense/bias:0']\n", "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n" ] } ], "source": [ "import tensorflow as tf\n", "from transformers import TFAutoModelForSequenceClassification\n", "\n", "#Hay dos categorías, así que ponemos 2 etiquetas (0 sano 1 tóxico)\n", "model = TFAutoModelForSequenceClassification.from_pretrained(\"dccuchile/bert-base-spanish-wwm-cased\", num_labels=2) " ] }, { "cell_type": "markdown", "id": "54d206b4", "metadata": {}, "source": [ "A hora vamos a convertir los datasets tokenizados en datasets de TensorFlow con el método .to_tf_dataset. Las entradas están en columns y la etiqueta en label_cols. El bach size es el número de ejemplos que se introducen en la red para que se entrene cada vez." ] }, { "cell_type": "code", "execution_count": 6, "id": "2ac843c2", "metadata": {}, "outputs": [], "source": [ "tf_train_dataset= train_dataset.to_tf_dataset(\n", "columns=[\"attention_mask\", \"input_ids\", \"token_type_ids\" ],\n", "label_cols=\"labels\",\n", "shuffle=True,\n", "collate_fn=data_collator,\n", "batch_size=8,\n", ")\n", "tf_validation_dataset= eval_dataset.to_tf_dataset(\n", "columns=[\"attention_mask\", \"input_ids\", \"token_type_ids\"],\n", "label_cols=\"labels\",\n", "shuffle=False,\n", "collate_fn=data_collator,\n", "batch_size=8,\n", ")\n" ] }, { "cell_type": "markdown", "id": "d07f651a", "metadata": {}, "source": [ "Compilamos" ] }, { "cell_type": "code", "execution_count": 7, "id": "72e85cab", "metadata": {}, "outputs": [], "source": [ "model.compile(\n", "optimizer=tf.keras.optimizers.Adam(learning_rate=5e-5),\n", "loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),\n", "metrics=tf.metrics.SparseCategoricalAccuracy(),\n", ")" ] }, { "cell_type": "markdown", "id": "f103f0de", "metadata": {}, "source": [ "## Cross-validation\n", "Se definen los parámetros de K-flod cross valdation en primer lugar. Al ser un dataset pequeño el nmero de \n", "splits será de 3." ] }, { "cell_type": "code", "execution_count": 8, "id": "924886a1", "metadata": {}, "outputs": [], "source": [ "from sklearn.model_selection import KFold\n", "from keras.callbacks import EarlyStopping\n", "num_splits = 3\n", "kf = KFold(num_splits, shuffle= True, random_state=42)\n" ] }, { "cell_type": "markdown", "id": "651afcdb", "metadata": {}, "source": [ "\n", "Ahora definimos el ciclo de validación cruzada" ] }, { "cell_type": "code", "execution_count": 9, "id": "f96f6bae", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Fold 1\n", "Epoch 1/10\n", "23/23 [==============================] - 497s 21s/step - loss: 0.6079 - sparse_categorical_accuracy: 0.6292 - val_loss: 0.4507 - val_sparse_categorical_accuracy: 0.8000\n", "Epoch 2/10\n", "23/23 [==============================] - 472s 21s/step - loss: 0.1810 - sparse_categorical_accuracy: 0.9551 - val_loss: 0.6494 - val_sparse_categorical_accuracy: 0.7556\n", "Epoch 3/10\n", "23/23 [==============================] - 476s 21s/step - loss: 0.0870 - sparse_categorical_accuracy: 0.9607 - val_loss: 0.6662 - val_sparse_categorical_accuracy: 0.7889\n", "Train\n", "Fold 1 - Loss: 0.19475573301315308, Accuracy: 0.949438214302063\n", "Val\n", "Fold 1 - Loss: 0.4506634473800659, Accuracy: 0.800000011920929\n", "Fold 2\n", "Epoch 1/10\n", "23/23 [==============================] - 466s 20s/step - loss: 0.3345 - sparse_categorical_accuracy: 0.8771 - val_loss: 0.1169 - val_sparse_categorical_accuracy: 0.9663\n", "Epoch 2/10\n", "23/23 [==============================] - 466s 20s/step - loss: 0.1234 - sparse_categorical_accuracy: 0.9721 - val_loss: 0.2614 - val_sparse_categorical_accuracy: 0.9438\n", "Epoch 3/10\n", "23/23 [==============================] - 467s 20s/step - loss: 0.1037 - sparse_categorical_accuracy: 0.9665 - val_loss: 0.2029 - val_sparse_categorical_accuracy: 0.9326\n", "Train\n", "Fold 2 - Loss: 0.08177345991134644, Accuracy: 0.9832402467727661\n", "Val\n", "Fold 2 - Loss: 0.11692894995212555, Accuracy: 0.966292142868042\n", "Fold 3\n", "Epoch 1/10\n", "23/23 [==============================] - 465s 20s/step - loss: 0.1484 - sparse_categorical_accuracy: 0.9665 - val_loss: 0.0247 - val_sparse_categorical_accuracy: 1.0000\n", "Epoch 2/10\n", "23/23 [==============================] - 464s 20s/step - loss: 0.0451 - sparse_categorical_accuracy: 0.9777 - val_loss: 0.0164 - val_sparse_categorical_accuracy: 1.0000\n", "Epoch 3/10\n", "23/23 [==============================] - 466s 20s/step - loss: 0.0066 - sparse_categorical_accuracy: 1.0000 - val_loss: 0.0064 - val_sparse_categorical_accuracy: 1.0000\n", "Epoch 4/10\n", "23/23 [==============================] - 463s 20s/step - loss: 0.0342 - sparse_categorical_accuracy: 0.9944 - val_loss: 0.0068 - val_sparse_categorical_accuracy: 1.0000\n", "Epoch 5/10\n", "23/23 [==============================] - 463s 20s/step - loss: 0.0441 - sparse_categorical_accuracy: 0.9888 - val_loss: 0.1121 - val_sparse_categorical_accuracy: 0.9551\n", "Train\n", "Fold 3 - Loss: 0.0020440793596208096, Accuracy: 1.0\n", "Val\n", "Fold 3 - Loss: 0.006410232279449701, Accuracy: 1.0\n" ] } ], "source": [ "#listas para almacenar las métricas en cada fold\n", "train_losses=[]\n", "train_accuracies=[]\n", "val_losses = []\n", "val_accuracies=[]\n", "\n", "for fold, (train_index, val_index) in enumerate(kf.split(train_dataset)):\n", " print (f\"Fold {fold + 1}\")\n", " \n", " #crear conjuntos de entrenamiento y validación para esta iteración\n", " train_fold_dataset = train_dataset.select(train_index)\n", " val_fold_dataset = train_dataset.select(val_index)\n", " \n", " #convertir los datasets a Tensorflow\n", " tf_train_fold_dataset= train_fold_dataset.to_tf_dataset(\n", " columns=[\"attention_mask\", \"input_ids\", \"token_type_ids\"],\n", " label_cols=\"labels\",\n", " shuffle=True,\n", " collate_fn=data_collator,\n", " batch_size=8,\n", " )\n", " \n", " tf_val_fold_dataset= val_fold_dataset.to_tf_dataset(\n", " columns=[\"attention_mask\", \"input_ids\", \"token_type_ids\"],\n", " label_cols=\"labels\",\n", " shuffle=False,\n", " collate_fn=data_collator,\n", " batch_size=8,\n", " )\n", " \n", " #early-stop\n", " early_stop=EarlyStopping(monitor=\"val_loss\",patience=2,mode=\"auto\", restore_best_weights=True)\n", " \n", " #entrenar el modelo \n", " model.fit(tf_train_fold_dataset, validation_data=tf_val_fold_dataset, epochs=10, callbacks=[early_stop])\n", " \n", " # Evaluar el modelo \n", " train_scores = model.evaluate(tf_train_fold_dataset, verbose=0)\n", " val_scores = model.evaluate(tf_val_fold_dataset, verbose=0)\n", " print(\"Train\")\n", " print(f\"Fold {fold + 1} - Loss: {train_scores[0]}, Accuracy: {train_scores[1]}\")\n", " print(\"Val\")\n", " print(f\"Fold {fold + 1} - Loss: {val_scores[0]}, Accuracy: {val_scores[1]}\")\n", " \n", " # Guardamos las cifras para después hacer la media\n", " train_losses.append(train_scores[0])\n", " train_accuracies.append(train_scores[1])\n", " val_losses.append(val_scores[0])\n", " val_accuracies.append(val_scores[1])\n", " \n", "\n" ] }, { "cell_type": "code", "execution_count": 10, "id": "4113ab57", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Mean Train Loss: 0.0928577574280401, Mean Train Accuracy: 0.977559487024943\n", "Mean Val Loss: 0.19133420987054706, Mean Val Accuracy: 0.922097384929657\n" ] } ], "source": [ "import numpy as np\n", "#Calcular las medidas de las métricas\n", "mean_train_loss = np.mean(train_losses)\n", "mean_train_accuracy = np.mean(train_accuracies)\n", "mean_val_loss = np.mean(val_losses)\n", "mean_val_accuracy = np. mean(val_accuracies)\n", "\n", "#Imprimir las medias de las métricas\n", "print(f\"Mean Train Loss: {mean_train_loss}, Mean Train Accuracy: {mean_train_accuracy}\")\n", "print(f\"Mean Val Loss: {mean_val_loss}, Mean Val Accuracy: {mean_val_accuracy}\")\n" ] }, { "cell_type": "code", "execution_count": null, "id": "0e0dff1a", "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "id": "ab3b230e", "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "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": "3.8.13" } }, "nbformat": 4, "nbformat_minor": 5 }