File size: 12,633 Bytes
4770791 |
1 2 3 4 5 6 7 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 |
{
"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 268 ejemplos de entrenamiento y 66 para testear."
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "b9a1f255",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'label': 1, 'text': 'Llamar muchas veces porque no te responden los mensajes'}"
]
},
"execution_count": 2,
"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'][201]"
]
},
{
"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. Par poder procesar el dataset en un solo paso, se utiliza el método dataset.map para preprocesar todo el dataset."
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "01673605",
"metadata": {},
"outputs": [],
"source": [
"from transformers import AutoTokenizer\n",
"\n",
"tokenizer = AutoTokenizer.from_pretrained(\"bert-base-multilingual-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": 4,
"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": 5,
"id": "698a98ca",
"metadata": {},
"outputs": [],
"source": [
"train_dataset = tokenized_datasets[\"train\"]\n",
"eval_dataset = tokenized_datasets[\"test\"]"
]
},
{
"cell_type": "markdown",
"id": "2c6d5142",
"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"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "40a05ad9",
"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": "38a6c521",
"metadata": {},
"source": [
"## Fine-tuning usando Fit\n",
"\n",
"En primer lugar, vamos a cargar el modelo TensorFlow con el número esperado e labels. En este caso, tenemos 2 categorías.\n",
"\n"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "843f218d",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"2023-08-09 20:57:30.270009: W tensorflow/tsl/framework/cpu_allocator_impl.cc:83] Allocation of 367248384 exceeds 10% of free system memory.\n",
"2023-08-09 20:57:30.504118: W tensorflow/tsl/framework/cpu_allocator_impl.cc:83] Allocation of 367248384 exceeds 10% of free system memory.\n",
"2023-08-09 20:57:30.551016: W tensorflow/tsl/framework/cpu_allocator_impl.cc:83] Allocation of 367248384 exceeds 10% of free system memory.\n",
"2023-08-09 20:57:37.098563: W tensorflow/tsl/framework/cpu_allocator_impl.cc:83] Allocation of 367248384 exceeds 10% of free system memory.\n",
"2023-08-09 20:57:37.939022: W tensorflow/tsl/framework/cpu_allocator_impl.cc:83] Allocation of 367248384 exceeds 10% of free system memory.\n",
"All PyTorch model weights were used when initializing TFBertForSequenceClassification.\n",
"\n",
"Some weights or buffers of the TF 2.0 model TFBertForSequenceClassification were not initialized from the PyTorch model and are newly initialized: ['classifier.weight', 'classifier.bias']\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(\"bert-base-multilingual-cased\", num_labels=2)"
]
},
{
"cell_type": "markdown",
"id": "a31780ca",
"metadata": {},
"source": [
"Ahora se aplica la función compile y fit como se haría con cualquier modelo Keras.\n",
"Compile configura la fase de entrenamiento del modelo antes comenzar a optimizar, por eso se elige el optimizador (en nuestro caso, Adam), la función de pérdida y las métricas que se usarań para evaluar el rendimiento que se han puesto en las celdas anteriores. \n",
"Fit entrena el modelo con los datos que se le han pasado, y al proporcionar un conjunto de validación se monitorea el rendimiento del modelo, por lo que se evalua mientras se entrena."
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "3e01c5fb",
"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",
")\n",
"\n"
]
},
{
"cell_type": "markdown",
"id": "4840d701",
"metadata": {},
"source": [
"sparse_categorical es el valor calculado en mi conjunto de datos de train, mientras que el que tiene el prefijo val es el que se calcula en el conjunto de datos de test. Si la métrica de test permanece igual o disminuye mientras aumenta el de train, el modelo está sobreajustando (overfitting)"
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "41f9da62",
"metadata": {},
"outputs": [],
"source": [
"from keras.callbacks import EarlyStopping\n",
"#en este modelo he observado overfitting, por lo que voy a utilizar Early stopping para detener el entrenamiento en el momento\n",
"#que se observe un incremento en el error de validación. \n",
"#Deja pasar 2 epochs antes de interrumpir el entrenamiento, quedándose con el mejor valor\n",
"early_stop=EarlyStopping(monitor=\"val_loss\",patience=2,mode=\"auto\", restore_best_weights=True)"
]
},
{
"cell_type": "markdown",
"id": "de7045d1",
"metadata": {},
"source": [
"Ahora aplicamos el fit con un epoch de 10, que es las veces que pasará cada prototipo por el entrenador, teniendo en cuenta que parará cuando el valor de loss empiece a detectar overfitting"
]
},
{
"cell_type": "code",
"execution_count": 11,
"id": "50e4097e",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Epoch 1/10\n",
"34/34 [==============================] - 592s 17s/step - loss: 0.7036 - sparse_categorical_accuracy: 0.4851 - val_loss: 0.6760 - val_sparse_categorical_accuracy: 0.6667\n",
"Epoch 2/10\n",
"34/34 [==============================] - 563s 17s/step - loss: 0.6176 - sparse_categorical_accuracy: 0.6866 - val_loss: 0.5916 - val_sparse_categorical_accuracy: 0.7121\n",
"Epoch 3/10\n",
"34/34 [==============================] - 562s 17s/step - loss: 0.6109 - sparse_categorical_accuracy: 0.6903 - val_loss: 0.7139 - val_sparse_categorical_accuracy: 0.5758\n",
"Epoch 4/10\n",
"34/34 [==============================] - 569s 17s/step - loss: 0.7128 - sparse_categorical_accuracy: 0.5858 - val_loss: 0.6790 - val_sparse_categorical_accuracy: 0.5606\n"
]
},
{
"data": {
"text/plain": [
"<keras.src.callbacks.History at 0x7fe030312700>"
]
},
"execution_count": 11,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"model.fit(tf_train_dataset, validation_data=tf_validation_dataset, epochs=10, callbacks=[early_stop])"
]
},
{
"cell_type": "code",
"execution_count": 12,
"id": "fbeef13e",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Model: \"tf_bert_for_sequence_classification\"\n",
"_________________________________________________________________\n",
" Layer (type) Output Shape Param # \n",
"=================================================================\n",
" bert (TFBertMainLayer) multiple 177853440 \n",
" \n",
" dropout_37 (Dropout) multiple 0 \n",
" \n",
" classifier (Dense) multiple 1538 \n",
" \n",
"=================================================================\n",
"Total params: 177854978 (678.46 MB)\n",
"Trainable params: 177854978 (678.46 MB)\n",
"Non-trainable params: 0 (0.00 Byte)\n",
"_________________________________________________________________\n"
]
}
],
"source": [
"model.summary()"
]
},
{
"cell_type": "markdown",
"id": "c4fa0fce",
"metadata": {},
"source": [
"Aunque aparecen durante el proceso de fit, imprimimos las cifras de loss y accuracy obtenidas del modelo.\n"
]
},
{
"cell_type": "code",
"execution_count": 14,
"id": "4113ab57",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"('loss', 0.5915697813034058)\n",
"('sparse_categorical_accuracy', 0.7121211886405945)\n"
]
}
],
"source": [
"\n",
"scores= model.evaluate(tf_validation_dataset, verbose=0)\n",
"print((model.metrics_names[0], scores[0]))\n",
"print((model.metrics_names[1], scores[1]))"
]
},
{
"cell_type": "markdown",
"id": "9e61a040",
"metadata": {},
"source": [
"# Guardando el modelo"
]
},
{
"cell_type": "markdown",
"id": "4af06209",
"metadata": {},
"source": [
"Para Guardarlo, utilizamos esl método save_model"
]
},
{
"cell_type": "code",
"execution_count": 22,
"id": "0e0dff1a",
"metadata": {},
"outputs": [],
"source": [
"model.save(\"BERTmULT-k-MMG.keras\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "3dfd1db6",
"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
}
|