修复Trainer fp16的bug; 添加使用中文词进行分类的例子

docs/source/_static/notebooks/文本分类.ipynb

@@ -47,7 +47,9 @@
   {
    "cell_type": "code",
    "execution_count": 1,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "from fastNLP.io import ChnSentiCorpLoader\n",
@@ -126,7 +128,9 @@
   {
    "cell_type": "code",
    "execution_count": 3,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "from fastNLP.io import ChnSentiCorpPipe\n",
@@ -280,7 +284,9 @@
   {
    "cell_type": "code",
    "execution_count": 9,
-   "metadata": {},
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
    "source": [
     "from torch import nn\n",
@@ -802,12 +808,222 @@
     "tester.test()"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 基于词进行文本分类"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "由于汉字中没有显示的字与字的边界，一般需要通过分词器先将句子进行分词操作。\n",
+    "下面的例子演示了如何不基于fastNLP已有的数据读取、预处理代码进行文本分类。"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### (1) 读取数据"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "这里我们继续以之前的数据为例，但这次我们不使用fastNLP自带的数据读取代码  "
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "from fastNLP.io import ChnSentiCorpLoader\n",
+    "\n",
+    "loader = ChnSentiCorpLoader()        # 初始化一个中文情感分类的loader\n",
+    "data_dir = loader.download()         # 这一行代码将自动下载数据到默认的缓存地址, 并将该地址返回"
+   ]
+  },
+  {
+   "cell_type": "markdown",
    "metadata": {},
+   "source": [
+    "下面我们先定义一个read_file_to_dataset的函数, 即给定一个文件路径，读取其中的内容，并返回一个DataSet。然后我们将所有的DataSet放入到DataBundle对象中来方便接下来的预处理"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "collapsed": true
+   },
    "outputs": [],
-   "source": []
+   "source": [
+    "import os\n",
+    "from fastNLP import DataSet, Instance\n",
+    "from fastNLP.io import DataBundle\n",
+    "\n",
+    "\n",
+    "def read_file_to_dataset(fp):\n",
+    "    ds = DataSet()\n",
+    "    with open(fp, 'r') as f:\n",
+    "        f.readline()  # 第一行是title名称，忽略掉\n",
+    "        for line in f:\n",
+    "            line = line.strip()\n",
+    "            target, chars = line.split('\\t')\n",
+    "            ins = Instance(target=target, raw_chars=chars)\n",
+    "            ds.append(ins)\n",
+    "    return ds\n",
+    "\n",
+    "data_bundle = DataBundle()\n",
+    "for name in ['train.tsv', 'dev.tsv', 'test.tsv']:\n",
+    "    fp = os.path.join(data_dir, name)\n",
+    "    ds = read_file_to_dataset(fp)\n",
+    "    data_bundle.set_dataset(name=name.split('.')[0], dataset=ds)\n",
+    "\n",
+    "print(data_bundle)  # 查看以下数据集的情况\n",
+    "# In total 3 datasets:\n",
+    "#    train has 9600 instances.\n",
+    "#    dev has 1200 instances.\n",
+    "#    test has 1200 instances."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### (2) 数据预处理"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "在这里，我们首先把句子通过 [fastHan](http://gitee.com/fastnlp/fastHan) 进行分词操作，然后创建词表，并将词语转换为序号。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "from fastHan import FastHan\n",
+    "from fastNLP import Vocabulary\n",
+    "\n",
+    "model=FastHan()\n",
+    "\n",
+    "# 定义分词处理操作\n",
+    "def word_seg(ins):\n",
+    "    raw_chars = ins['raw_chars']\n",
+    "    # 由于有些句子比较长，我们只截取前128个汉字\n",
+    "    raw_words = model(raw_chars[:128], target='CWS')[0]\n",
+    "    return raw_words\n",
+    "\n",
+    "for name, ds in data_bundle.iter_datasets():\n",
+    "    # apply函数将对内部的instance依次执行word_seg操作，并把其返回值放入到raw_words这个field\n",
+    "    ds.apply(word_seg, new_field_name='raw_words')\n",
+    "    # 除了apply函数，fastNLP还支持apply_field, apply_more(可同时创建多个field)等操作\n",
+    "\n",
+    "vocab = Vocabulary()\n",
+    "\n",
+    "# 对raw_words列创建词表, 建议把非训练集的dataset放在no_create_entry_dataset参数中\n",
+    "# 也可以通过add_word(), add_word_lst()等建立词表，请参考http://www.fastnlp.top/docs/fastNLP/tutorials/tutorial_2_vocabulary.html\n",
+    "vocab.from_dataset(data_bundle.get_dataset('train'), field_name='raw_words', \n",
+    "                no_create_entry_dataset=[data_bundle.get_dataset('dev'), \n",
+    "                                            data_bundle.get_dataset('test')]) \n",
+    "\n",
+    "# 将建立好词表的Vocabulary用于对raw_words列建立词表，并把转为序号的列存入到words列\n",
+    "vocab.index_dataset(data_bundle.get_dataset('train'), data_bundle.get_dataset('dev'), \n",
+    "                data_bundle.get_dataset('test'), field_name='raw_words', new_field_name='words')\n",
+    "\n",
+    "# 建立target的词表，target的词表一般不需要padding和unknown\n",
+    "target_vocab = Vocabulary(padding=None, unknown=None) \n",
+    "# 一般情况下我们可以只用训练集建立target的词表\n",
+    "target_vocab.from_dataset(data_bundle.get_dataset('train'), field_name='target') \n",
+    "# 如果没有传递new_field_name, 则默认覆盖原词表\n",
+    "target_vocab.index_dataset(data_bundle.get_dataset('train'), data_bundle.get_dataset('dev'), \n",
+    "                data_bundle.get_dataset('test'), field_name='target')\n",
+    "\n",
+    "# 我们可以把词表保存到data_bundle中，方便之后使用\n",
+    "data_bundle.set_vocab(field_name='words', vocab=vocab)\n",
+    "data_bundle.set_vocab(field_name='target', vocab=target_vocab)\n",
+    "\n",
+    "# 我们把words和target分别设置为input和target，这样它们才会在训练循环中被取出并自动padding, 有关这部分更多的内容参考\n",
+    "#  http://www.fastnlp.top/docs/fastNLP/tutorials/tutorial_6_datasetiter.html\n",
+    "data_bundle.set_target('target')\n",
+    "data_bundle.set_input('words')  # DataSet也有这两个接口\n",
+    "# 如果某些field，您希望它被设置为target或者input，但是不希望fastNLP自动padding或需要使用特定的padding方式，请参考\n",
+    "#  http://www.fastnlp.top/docs/fastNLP/fastNLP.core.dataset.html\n",
+    "\n",
+    "print(data_bundle.get_dataset('train')[:2])  # 我们可以看一下当前dataset的内容"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### (3) 选择预训练词向量"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "这里我们选择腾讯的预训练中文词向量，可以在 [腾讯词向量](https://ai.tencent.com/ailab/nlp/en/embedding.html) 处下载并解压。这里我们不能直接使用BERT，因为BERT是基于中文字进行预训练的。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "from fastNLP.embeddings import StaticEmbedding\n",
+    "\n",
+    "word2vec_embed = StaticEmbedding(data_bundle.get_vocab('words'), \n",
+    "                                 model_dir_or_name='/path/to/Tencent_AILab_ChineseEmbedding.txt')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "# 初始化模型\n",
+    "model = BiLSTMMaxPoolCls(word2vec_embed, len(data_bundle.get_vocab('target')))\n",
+    "\n",
+    "# 开始训练\n",
+    "loss = CrossEntropyLoss()\n",
+    "optimizer = Adam(model.parameters(), lr=0.001)\n",
+    "metric = AccuracyMetric()\n",
+    "device = 0 if torch.cuda.is_available() else 'cpu'  # 如果有gpu的话在gpu上运行，训练速度会更快\n",
+    "\n",
+    "trainer = Trainer(train_data=data_bundle.get_dataset('train'), model=model, loss=loss, \n",
+    "                  optimizer=optimizer, batch_size=32, dev_data=data_bundle.get_dataset('dev'),\n",
+    "                  metrics=metric, device=device)\n",
+    "trainer.train()  # 开始训练，训练完成之后默认会加载在dev上表现最好的模型\n",
+    "\n",
+    "# 在测试集上测试一下模型的性能\n",
+    "from fastNLP import Tester\n",
+    "print(\"Performance on test is:\")\n",
+    "tester = Tester(data=data_bundle.get_dataset('test'), model=model, metrics=metric, batch_size=64, device=device)\n",
+    "tester.test()"
+   ]
   }
  ],
  "metadata": {
@@ -826,7 +1042,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.6.7"
+   "version": "3.6.10"
   }
  },
  "nbformat": 4,