{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 使用 Trainer API 或者 Keras 微调一个模型"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Install the Transformers, Datasets, and Evaluate libraries to run this notebook."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install datasets evaluate transformers[sentencepiece]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from datasets import load_dataset\n",
    "from transformers import AutoTokenizer, DataCollatorWithPadding\n",
    "import numpy as np\n",
    "\n",
    "raw_datasets = load_dataset(\"glue\", \"mrpc\")\n",
    "checkpoint = \"bert-base-uncased\"\n",
    "tokenizer = AutoTokenizer.from_pretrained(checkpoint)\n",
    "\n",
    "\n",
    "def tokenize_function(example):\n",
    "    return tokenizer(example[\"sentence1\"], example[\"sentence2\"], truncation=True)\n",
    "\n",
    "\n",
    "tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)\n",
    "\n",
    "data_collator = DataCollatorWithPadding(tokenizer=tokenizer, return_tensors=\"tf\")\n",
    "\n",
    "tf_train_dataset = tokenized_datasets[\"train\"].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_validation_dataset = tokenized_datasets[\"validation\"].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",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import TFAutoModelForSequenceClassification\n",
    "\n",
    "model = TFAutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tensorflow.keras.losses import SparseCategoricalCrossentropy\n",
    "\n",
    "model.compile(\n",
    "    optimizer=\"adam\",\n",
    "    loss=SparseCategoricalCrossentropy(from_logits=True),\n",
    "    metrics=[\"accuracy\"],\n",
    ")\n",
    "model.fit(\n",
    "    tf_train_dataset,\n",
    "    validation_data=tf_validation_dataset,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tensorflow.keras.optimizers.schedules import PolynomialDecay\n",
    "\n",
    "batch_size = 8\n",
    "num_epochs = 3\n",
    "# 训练步数是数据集中的样本数除以batch size再乘以 epoch。\n",
    "# 注意这里的tf_train_dataset是一个转化为batch后的 tf.data.Dataset，\n",
    "# 不是原来的 Hugging Face Dataset，所以它的 len() 已经是 num_samples // batch_size。\n",
    "num_train_steps = len(tf_train_dataset) * num_epochs\n",
    "lr_scheduler = PolynomialDecay(\n",
    "    initial_learning_rate=5e-5, end_learning_rate=0.0, decay_steps=num_train_steps\n",
    ")\n",
    "from tensorflow.keras.optimizers import Adam\n",
    "\n",
    "opt = Adam(learning_rate=lr_scheduler)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "\n",
    "model = TFAutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)\n",
    "loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\n",
    "model.compile(optimizer=opt, loss=loss, metrics=[\"accuracy\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.fit(tf_train_dataset, validation_data=tf_validation_dataset, epochs=3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "preds = model.predict(tf_validation_dataset)[\"logits\"]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(408, 2) (408,)"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "class_preds = np.argmax(preds, axis=1)\n",
    "print(preds.shape, class_preds.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'accuracy': 0.8578431372549019, 'f1': 0.8996539792387542}"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import evaluate\n",
    "\n",
    "metric = evaluate.load(\"glue\", \"mrpc\")\n",
    "metric.compute(predictions=class_preds, references=raw_datasets[\"validation\"][\"label\"])"
   ]
  }
 ],
 "metadata": {
  "colab": {
   "name": "使用 Trainer API 或者 Keras 微调一个模型",
   "provenance": []
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}