{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "\n", "## 用N-Gram模型在莎士比亚文集中训练word embedding\n", "N-gram 是计算机语言学和概率论范畴内的概念，是指给定的一段文本中N个项目的序列。\n", "N=1 时 N-gram 又称为 unigram，N=2 称为 bigram，N=3 称为 trigram，以此类推。实际应用通常采用 bigram 和 trigram 进行计算。\n", "本示例在莎士比亚文集上实现了trigram。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# 环境\n", "本教程基于paddle-develop编写，如果您的环境不是本版本，请先安装paddle-develop。" ] }, { "cell_type": "code", "execution_count": 17, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "'0.0.0'" ] }, "execution_count": 17, "metadata": {}, "output_type": "execute_result" } ], "source": [ "import paddle\n", "paddle.__version__" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 数据集&&相关参数\n", "训练数据集采用了莎士比亚文集，[下载](https://ocw.mit.edu/ans7870/6/6.006/s08/lecturenotes/files/t8.shakespeare.txt)，保存为txt格式即可。
\n", "context_size设为2，意味着是trigram。embedding_dim设为256。" ] }, { "cell_type": "code", "execution_count": 18, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "--2020-09-03 08:41:10-- https://ocw.mit.edu/ans7870/6/6.006/s08/lecturenotes/files/t8.shakespeare.txt\n", "正在解析主机 ocw.mit.edu (ocw.mit.edu)... 151.101.230.133\n", "正在连接 ocw.mit.edu (ocw.mit.edu)|151.101.230.133|:443... 已连接。\n", "已发出 HTTP 请求，正在等待回应... 200 OK\n", "长度：5458199 (5.2M) [text/plain]\n", "正在保存至: “t8.shakespeare.txt.1”\n", "\n", "t8.shakespeare.txt. 100%[===================>] 5.21M 26.1KB/s 用时 4m 14s \n", "\n", "2020-09-03 08:45:25 (21.0 KB/s) - 已保存 “t8.shakespeare.txt.1” [5458199/5458199])\n", "\n" ] } ], "source": [ "!wget https://ocw.mit.edu/ans7870/6/6.006/s08/lecturenotes/files/t8.shakespeare.txt" ] }, { "cell_type": "code", "execution_count": 19, "metadata": {}, "outputs": [], "source": [ "embedding_dim = 256\n", "context_size = 2" ] }, { "cell_type": "code", "execution_count": 20, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Length of text: 5458199 characters\n" ] } ], "source": [ "# 文件路径\n", "path_to_file = './t8.shakespeare.txt'\n", "test_sentence = open(path_to_file, 'rb').read().decode(encoding='utf-8')\n", "\n", "# 文本长度是指文本中的字符个数\n", "print ('Length of text: {} characters'.format(len(test_sentence)))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 去除标点符号\n", "用`string`库中的punctuation，完成英文符号的替换。" ] }, { "cell_type": "code", "execution_count": 21, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "{'!': '', '\"': '', '#': '', '$': '', '%': '', '&': '', \"'\": '', '(': '', ')': '', '*': '', '+': '', ',': '', '-': '', '.': '', '/': '', ':': '', ';': '', '<': '', '=': '', '>': '', '?': '', '@': '', '[': '', '\\\\': '', ']': '', '^': '', '_': '', '`': '', '{': '', '|': '', '}': '', '~': ''}\n" ] } ], "source": [ "from string import punctuation\n", "process_dicts={i:'' for i in punctuation}\n", "print(process_dicts)" ] }, { "cell_type": "code", "execution_count": 22, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "28343\n" ] } ], "source": [ "punc_table = str.maketrans(process_dicts)\n", "test_sentence = test_sentence.translate(punc_table)\n", "test_sentence = test_sentence.lower().split()\n", "vocab = set(test_sentence)\n", "print(len(vocab))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 数据预处理\n", "将文本被拆成了元组的形式，格式为(('第一个词', '第二个词'), '第三个词');其中，第三个词就是我们的目标。" ] }, { "cell_type": "code", "execution_count": 23, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "[[['this', 'is'], 'the'], [['is', 'the'], '100th'], [['the', '100th'], 'etext']]\n" ] } ], "source": [ "trigram = [[[test_sentence[i], test_sentence[i + 1]], test_sentence[i + 2]]\n", " for i in range(len(test_sentence) - 2)]\n", "\n", "word_to_idx = {word: i for i, word in enumerate(vocab)}\n", "idx_to_word = {word_to_idx[word]: word for word in word_to_idx}\n", "# 看一下数据集\n", "print(trigram[:3])" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 构建`Dataset`类加载数据\n", "用`paddle.io.Dataset`构建数据集，然后作为参数传入到`paddle.io.DataLoader`，完成数据集的加载。" ] }, { "cell_type": "code", "execution_count": 24, "metadata": {}, "outputs": [], "source": [ "import paddle\n", "import numpy as np\n", "batch_size = 256\n", "paddle.disable_static()\n", "class TrainDataset(paddle.io.Dataset):\n", " def __init__(self, tuple_data):\n", " self.tuple_data = tuple_data\n", "\n", " def __getitem__(self, idx):\n", " data = self.tuple_data[idx][0]\n", " label = self.tuple_data[idx][1]\n", " data = np.array(list(map(lambda w: word_to_idx[w], data)))\n", " label = np.array(word_to_idx[label])\n", " return data, label\n", " \n", " def __len__(self):\n", " return len(self.tuple_data)\n", "train_dataset = TrainDataset(trigram)\n", "train_loader = paddle.io.DataLoader(train_dataset,places=paddle.fluid.cpu_places(), return_list=True,\n", " shuffle=True, batch_size=batch_size, drop_last=True)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 组网&训练\n", "这里用paddle动态图的方式组网。为了构建Trigram模型，用一层 `Embedding` 与两层 `Linear` 完成构建。`Embedding` 层对输入的前两个单词embedding，然后输入到后面的两个`Linear`层中，完成特征提取。" ] }, { "cell_type": "code", "execution_count": 25, "metadata": {}, "outputs": [], "source": [ "import paddle\n", "import numpy as np\n", "hidden_size = 1024\n", "class NGramModel(paddle.nn.Layer):\n", " def __init__(self, vocab_size, embedding_dim, context_size):\n", " super(NGramModel, self).__init__()\n", " self.embedding = paddle.nn.Embedding(size=[vocab_size, embedding_dim])\n", " self.linear1 = paddle.nn.Linear(context_size * embedding_dim, hidden_size)\n", " self.linear2 = paddle.nn.Linear(hidden_size, len(vocab))\n", "\n", " def forward(self, x):\n", " x = self.embedding(x)\n", " x = paddle.reshape(x, [-1, context_size * embedding_dim])\n", " x = self.linear1(x)\n", " x = paddle.nn.functional.relu(x)\n", " x = self.linear2(x)\n", " return x" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 定义`train()`函数，对模型进行训练。" ] }, { "cell_type": "code", "execution_count": 28, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "epoch: 0, batch_id: 0, loss is: [10.252256]\n", "epoch: 0, batch_id: 100, loss is: [7.0485706]\n", "epoch: 0, batch_id: 200, loss is: [7.282592]\n", "epoch: 0, batch_id: 300, loss is: [6.9604626]\n", "epoch: 0, batch_id: 400, loss is: [6.7308316]\n", "epoch: 0, batch_id: 500, loss is: [6.7940483]\n", "epoch: 0, batch_id: 600, loss is: [6.6574802]\n", "epoch: 0, batch_id: 700, loss is: [6.862562]\n", "epoch: 0, batch_id: 800, loss is: [7.2091002]\n", "epoch: 0, batch_id: 900, loss is: [7.0172606]\n", "epoch: 0, batch_id: 1000, loss is: [6.9888105]\n", "epoch: 0, batch_id: 1100, loss is: [6.9609995]\n", "epoch: 0, batch_id: 1200, loss is: [6.550024]\n", "epoch: 0, batch_id: 1300, loss is: [6.714109]\n", "epoch: 0, batch_id: 1400, loss is: [6.995716]\n", "epoch: 0, batch_id: 1500, loss is: [6.939434]\n", "epoch: 0, batch_id: 1600, loss is: [6.5966253]\n", "epoch: 0, batch_id: 1700, loss is: [6.9880104]\n", "epoch: 0, batch_id: 1800, loss is: [6.6459093]\n", "epoch: 0, batch_id: 1900, loss is: [6.8095036]\n", "epoch: 0, batch_id: 2000, loss is: [6.8447037]\n", "epoch: 0, batch_id: 2100, loss is: [6.8313]\n", "epoch: 0, batch_id: 2200, loss is: [6.808483]\n", "epoch: 0, batch_id: 2300, loss is: [6.502908]\n", "epoch: 0, batch_id: 2400, loss is: [6.561283]\n", "epoch: 0, batch_id: 2500, loss is: [7.0093765]\n", "epoch: 0, batch_id: 2600, loss is: [6.512396]\n", "epoch: 0, batch_id: 2700, loss is: [6.809763]\n", "epoch: 0, batch_id: 2800, loss is: [6.806659]\n", "epoch: 0, batch_id: 2900, loss is: [6.95402]\n", "epoch: 0, batch_id: 3000, loss is: [6.634927]\n", "epoch: 0, batch_id: 3100, loss is: [6.644098]\n", "epoch: 0, batch_id: 3200, loss is: [6.705504]\n", "epoch: 0, batch_id: 3300, loss is: [6.2121572]\n", "epoch: 0, batch_id: 3400, loss is: [6.638401]\n", "epoch: 0, batch_id: 3500, loss is: [6.986831]\n", "epoch: 1, batch_id: 0, loss is: [6.795429]\n", "epoch: 1, batch_id: 100, loss is: [6.582568]\n", "epoch: 1, batch_id: 200, loss is: [6.527663]\n", "epoch: 1, batch_id: 300, loss is: [6.714637]\n", "epoch: 1, batch_id: 400, loss is: [6.574902]\n", "epoch: 1, batch_id: 500, loss is: [6.305031]\n", "epoch: 1, batch_id: 600, loss is: [6.803609]\n", "epoch: 1, batch_id: 700, loss is: [6.2429113]\n", "epoch: 1, batch_id: 800, loss is: [6.7452283]\n", "epoch: 1, batch_id: 900, loss is: [6.383783]\n", "epoch: 1, batch_id: 1000, loss is: [6.4906135]\n", "epoch: 1, batch_id: 1100, loss is: [6.6007314]\n", "epoch: 1, batch_id: 1200, loss is: [6.63466]\n", "epoch: 1, batch_id: 1300, loss is: [6.540749]\n", "epoch: 1, batch_id: 1400, loss is: [6.7752547]\n", "epoch: 1, batch_id: 1500, loss is: [6.2411666]\n", "epoch: 1, batch_id: 1600, loss is: [6.540929]\n", "epoch: 1, batch_id: 1700, loss is: [6.6563463]\n", "epoch: 1, batch_id: 1800, loss is: [6.4592104]\n", "epoch: 1, batch_id: 1900, loss is: [7.0268345]\n", "epoch: 1, batch_id: 2000, loss is: [6.803793]\n", "epoch: 1, batch_id: 2100, loss is: [6.8454733]\n", "epoch: 1, batch_id: 2200, loss is: [6.651756]\n", "epoch: 1, batch_id: 2300, loss is: [6.5876465]\n", "epoch: 1, batch_id: 2400, loss is: [6.258934]\n", "epoch: 1, batch_id: 2500, loss is: [6.5422425]\n", "epoch: 1, batch_id: 2600, loss is: [6.184501]\n", "epoch: 1, batch_id: 2700, loss is: [6.6847773]\n", "epoch: 1, batch_id: 2800, loss is: [6.684101]\n", "epoch: 1, batch_id: 2900, loss is: [6.374978]\n", "epoch: 1, batch_id: 3000, loss is: [6.8277273]\n", "epoch: 1, batch_id: 3100, loss is: [6.5195084]\n", "epoch: 1, batch_id: 3200, loss is: [6.311832]\n", "epoch: 1, batch_id: 3300, loss is: [6.4282994]\n", "epoch: 1, batch_id: 3400, loss is: [6.603338]\n", "epoch: 1, batch_id: 3500, loss is: [6.4541807]\n" ] } ], "source": [ "import time\n", "vocab_size = len(vocab)\n", "epochs = 2\n", "losses = []\n", "def train(model):\n", " model.train()\n", " optim = paddle.optimizer.Adam(learning_rate=0.01, parameters=model.parameters())\n", " for epoch in range(epochs):\n", " for batch_id, data in enumerate(train_loader()):\n", " x_data = data[0]\n", " y_data = data[1]\n", " predicts = model(x_data)\n", " y_data = paddle.nn.functional.one_hot(y_data, len(vocab))\n", " loss = paddle.nn.functional.softmax_with_cross_entropy(predicts, y_data,soft_label=True)\n", " avg_loss = paddle.mean(loss)\n", " avg_loss.backward()\n", " if batch_id % 100 == 0:\n", " losses.append(avg_loss.numpy())\n", " print(\"epoch: {}, batch_id: {}, loss is: {}\".format(epoch, batch_id, avg_loss.numpy())) \n", " optim.minimize(avg_loss)\n", " model.clear_gradients()\n", "model = NGramModel(vocab_size, embedding_dim, context_size)\n", "train(model)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 打印loss下降曲线\n", "通过可视化loss的曲线，可以看到模型训练的效果。" ] }, { "cell_type": "code", "execution_count": 29, "metadata": {}, "outputs": [ { "data": { "text/plain": [ "[]" ] }, "execution_count": 29, "metadata": {}, "output_type": "execute_result" }, { "data": { "image/png": "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\n", "text/plain": [ "

" ] }, "metadata": { "needs_background": "light" }, "output_type": "display_data" } ], "source": [ "import matplotlib.pyplot as plt\n", "import matplotlib.ticker as ticker\n", "%matplotlib inline\n", "\n", "plt.figure()\n", "plt.plot(losses)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 预测\n", "用训练好的模型进行预测。" ] }, { "cell_type": "code", "execution_count": 30, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "the input words is: of, william\n", "the predict words is: shakespeare\n", "the true words is: shakespeare\n" ] }, { "name": "stderr", "output_type": "stream", "text": [ "/Library/Python/3.7/site-packages/ipykernel/ipkernel.py:287: DeprecationWarning: `should_run_async` will not call `transform_cell` automatically in the future. Please pass the result to `transformed_cell` argument and any exception that happen during thetransform in `preprocessing_exc_tuple` in IPython 7.17 and above.\n", " and should_run_async(code)\n" ] } ], "source": [ "import random\n", "def test(model):\n", " model.eval()\n", " # 从最后10组数据中随机选取1个\n", " idx = random.randint(len(trigram)-10, len(trigram)-1)\n", " print('the input words is: ' + trigram[idx][0][0] + ', ' + trigram[idx][0][1])\n", " x_data = list(map(lambda w: word_to_idx[w], trigram[idx][0]))\n", " x_data = paddle.to_tensor(np.array(x_data))\n", " predicts = model(x_data)\n", " predicts = predicts.numpy().tolist()[0]\n", " predicts = predicts.index(max(predicts))\n", " print('the predict words is: ' + idx_to_word[predicts])\n", " y_data = trigram[idx][1]\n", " print('the true words is: ' + y_data)\n", "test(model)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" } }, "nbformat": 4, "nbformat_minor": 4 }