add code 6.4

code/chapter06_RNN/6.4_rnn-scratch.ipynb

+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "0.4.1\n",
+      "cpu\n"
+     ]
+    }
+   ],
+   "source": [
+    "import time\n",
+    "import math\n",
+    "import numpy as np\n",
+    "import torch\n",
+    "from torch import nn, optim\n",
+    "import torch.nn.functional as F\n",
+    "\n",
+    "import sys\n",
+    "sys.path.append(\"..\") \n",
+    "import d2lzh_pytorch as d2l\n",
+    "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
+    "\n",
+    "print(torch.__version__)\n",
+    "print(device)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "(corpus_indices, char_to_idx, idx_to_char, vocab_size) = d2l.load_data_jay_lyrics()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "tensor([[1., 0., 0.,  ..., 0., 0., 0.],\n",
+       "        [0., 0., 1.,  ..., 0., 0., 0.]])"
+      ]
+     },
+     "execution_count": 3,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "def one_hot(x, n_class, dtype=torch.float32): \n",
+    "    # X shape: (batch), output shape: (batch, n_class)\n",
+    "    x = x.long()\n",
+    "    res = torch.zeros(x.shape[0], n_class, dtype=dtype)\n",
+    "    res.scatter_(1, x.view(-1, 1), 1)\n",
+    "    return res\n",
+    "    \n",
+    "x = torch.tensor([0, 2])\n",
+    "one_hot(x, vocab_size)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "5 torch.Size([2, 1027])\n"
+     ]
+    }
+   ],
+   "source": [
+    "def to_onehot(X, n_class):  # 本函数已保存在d2lzh包中方便以后使用\n",
+    "    # X shape: (batch, seq_len), output: seq_len elements of (batch, n_class)\n",
+    "    return [one_hot(X[:, i], n_class) for i in range(X.shape[1])]\n",
+    "\n",
+    "X = torch.arange(10).view(2, 5)\n",
+    "inputs = to_onehot(X, vocab_size)\n",
+    "print(len(inputs), inputs[0].shape)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "will use cpu\n"
+     ]
+    }
+   ],
+   "source": [
+    "num_inputs, num_hiddens, num_outputs = vocab_size, 256, vocab_size\n",
+    "print('will use', device)\n",
+    "\n",
+    "def get_params():\n",
+    "    def _one(shape):\n",
+    "        ts = torch.tensor(np.random.normal(0, 0.01, size=shape), device=device, dtype=torch.float32)\n",
+    "        return torch.nn.Parameter(ts, requires_grad=True)\n",
+    "\n",
+    "    # 隐藏层参数\n",
+    "    W_xh = _one((num_inputs, num_hiddens))\n",
+    "    W_hh = _one((num_hiddens, num_hiddens))\n",
+    "    b_h = torch.nn.Parameter(torch.zeros(num_hiddens, device=device, requires_grad=True))\n",
+    "    # 输出层参数\n",
+    "    W_hq = _one((num_hiddens, num_outputs))\n",
+    "    b_q = torch.nn.Parameter(torch.zeros(num_outputs, device=device, requires_grad=True))\n",
+    "    return nn.ParameterList([W_xh, W_hh, b_h, W_hq, b_q])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "def init_rnn_state(batch_size, num_hiddens, device):\n",
+    "    return (torch.zeros((batch_size, num_hiddens), device=device), )"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "def rnn(inputs, state, params):\n",
+    "    # inputs和outputs皆为num_steps个形状为(batch_size, vocab_size)的矩阵\n",
+    "    W_xh, W_hh, b_h, W_hq, b_q = params\n",
+    "    H, = state\n",
+    "    outputs = []\n",
+    "    for X in inputs:\n",
+    "        H = torch.tanh(torch.matmul(X, W_xh) + torch.matmul(H, W_hh) + b_h)\n",
+    "        Y = torch.matmul(H, W_hq) + b_q\n",
+    "        outputs.append(Y)\n",
+    "    return outputs, (H,)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "5 torch.Size([2, 1027]) torch.Size([2, 256])\n"
+     ]
+    }
+   ],
+   "source": [
+    "state = init_rnn_state(X.shape[0], num_hiddens, device)\n",
+    "inputs = to_onehot(X.to(device), vocab_size)\n",
+    "params = get_params()\n",
+    "outputs, state_new = rnn(inputs, state, params)\n",
+    "print(len(outputs), outputs[0].shape, state_new[0].shape)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "# 本函数已保存在d2lzh_pytorch包中方便以后使用\n",
+    "def predict_rnn(prefix, num_chars, rnn, params, init_rnn_state,\n",
+    "                num_hiddens, vocab_size, device, idx_to_char, char_to_idx):\n",
+    "    state = init_rnn_state(1, num_hiddens, device)\n",
+    "    output = [char_to_idx[prefix[0]]]\n",
+    "    for t in range(num_chars + len(prefix) - 1):\n",
+    "        # 将上一时间步的输出作为当前时间步的输入\n",
+    "        X = to_onehot(torch.tensor([[output[-1]]], device=device), vocab_size)\n",
+    "        # 计算输出和更新隐藏状态\n",
+    "        (Y, state) = rnn(X, state, params)\n",
+    "        # 下一个时间步的输入是prefix里的字符或者当前的最佳预测字符\n",
+    "        if t < len(prefix) - 1:\n",
+    "            output.append(char_to_idx[prefix[t + 1]])\n",
+    "        else:\n",
+    "            output.append(int(Y[0].argmax(dim=1).item()))\n",
+    "    return ''.join([idx_to_char[i] for i in output])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'分开爽只忆干走抄蝴配碑鹰'"
+      ]
+     },
+     "execution_count": 10,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "predict_rnn('分开', 10, rnn, params, init_rnn_state, num_hiddens, vocab_size,\n",
+    "            device, idx_to_char, char_to_idx)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 36,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "# 本函数已保存在d2lzh_pytorch包中方便以后使用\n",
+    "def grad_clipping(params, theta, device):\n",
+    "    norm = torch.tensor([0.0], device=device)\n",
+    "    for param in params:\n",
+    "        norm += (param.grad.data ** 2).sum()\n",
+    "    norm = norm.sqrt().item()\n",
+    "    if norm > theta:\n",
+    "        for param in params:\n",
+    "            param.grad.data *= (theta / norm)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 40,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "# 本函数已保存在d2lzh包中方便以后使用\n",
+    "def train_and_predict_rnn(rnn, get_params, init_rnn_state, num_hiddens,\n",
+    "                          vocab_size, device, corpus_indices, idx_to_char,\n",
+    "                          char_to_idx, is_random_iter, num_epochs, num_steps,\n",
+    "                          lr, clipping_theta, batch_size, pred_period,\n",
+    "                          pred_len, prefixes):\n",
+    "    if is_random_iter:\n",
+    "        data_iter_fn = d2l.data_iter_random\n",
+    "    else:\n",
+    "        data_iter_fn = d2l.data_iter_consecutive\n",
+    "    params = get_params()\n",
+    "    loss = nn.CrossEntropyLoss()\n",
+    "\n",
+    "    for epoch in range(num_epochs):\n",
+    "        if not is_random_iter:  # 如使用相邻采样，在epoch开始时初始化隐藏状态\n",
+    "            state = init_rnn_state(batch_size, num_hiddens, device)\n",
+    "        l_sum, n, start = 0.0, 0, time.time()\n",
+    "        data_iter = data_iter_fn(corpus_indices, batch_size, num_steps, device)\n",
+    "        for X, Y in data_iter:\n",
+    "            if is_random_iter:  # 如使用随机采样，在每个小批量更新前初始化隐藏状态\n",
+    "                state = init_rnn_state(batch_size, num_hiddens, device)\n",
+    "            else:  # 否则需要使用detach函数从计算图分离隐藏状态\n",
+    "                for s in state:\n",
+    "                    s.detach_()\n",
+    "            \n",
+    "            inputs = to_onehot(X, vocab_size)\n",
+    "            # outputs有num_steps个形状为(batch_size, vocab_size)的矩阵\n",
+    "            (outputs, state) = rnn(inputs, state, params)\n",
+    "            # 拼接之后形状为(num_steps * batch_size, vocab_size)\n",
+    "            outputs = torch.cat(outputs, dim=0)\n",
+    "            # Y的形状是(batch_size, num_steps)，转置后再变成长度为\n",
+    "            # batch * num_steps 的向量，这样跟输出的行一一对应\n",
+    "            y = torch.transpose(Y, 0, 1).flatten()\n",
+    "            # 使用交叉熵损失计算平均分类误差\n",
+    "            l = loss(outputs, y.long())\n",
+    "            \n",
+    "            # 梯度清0\n",
+    "            if params[0].grad is not None:\n",
+    "                for param in params:\n",
+    "                    param.grad.data.zero_()\n",
+    "            l.backward()\n",
+    "            grad_clipping(params, clipping_theta, device)  # 裁剪梯度\n",
+    "            d2l.sgd(params, lr, 1)  # 因为误差已经取过均值，梯度不用再做平均\n",
+    "            l_sum += l.item() * y.shape[0]\n",
+    "            n += y.shape[0]\n",
+    "\n",
+    "        if (epoch + 1) % pred_period == 0:\n",
+    "            print('epoch %d, perplexity %f, time %.2f sec' % (\n",
+    "                epoch + 1, math.exp(l_sum / n), time.time() - start))\n",
+    "            for prefix in prefixes:\n",
+    "                print(' -', predict_rnn(prefix, pred_len, rnn, params, init_rnn_state,\n",
+    "                    num_hiddens, vocab_size, device, idx_to_char, char_to_idx))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 41,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": [
+    "num_epochs, num_steps, batch_size, lr, clipping_theta = 250, 35, 32, 1e2, 1e-2\n",
+    "pred_period, pred_len, prefixes = 50, 50, ['分开', '不分开']"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 42,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "epoch 50, perplexity 69.368188, time 0.70 sec\n",
+      " - 分开 我不要的可爱女人 坏坏的让我疯狂的可爱女人 坏坏的让我疯狂的可爱女人 坏坏的让我疯狂的可爱女人 坏\n",
+      " - 不分开 快颗的双 我不定空 我有一场的溪 一知哈觉 我不要 别怪两 我给就的可爱女人 坏坏的让我疯狂的可爱\n",
+      "epoch 100, perplexity 10.391687, time 0.57 sec\n",
+      " - 分开 一颗在双截棍 哼哼哈 一直两 我想就这样牵着你的手不放开 爱能不能够永远单纯没有 看星形 一颗两颗\n",
+      " - 不分开吗 我不要再想 我不能再想 我不能再想 我不能再想 我不能再想 我不能再想 我不能再想 我不能再想 \n",
+      "epoch 150, perplexity 2.892629, time 0.58 sec\n",
+      " - 分开 一颗会 干什么 我满就这手牵着你的手不放开 爱能不能够永远单纯没有悲哀 我 想带你骑单车 我 想和\n",
+      " - 不分开吗 我后能爸 你知我妈 这样了直不屈 一身正气 快使用双截棍 哼哼哈兮 快使用双截棍 哼哼哈兮 快使\n",
+      "epoch 200, perplexity 1.587798, time 0.59 sec\n",
+      " - 分开 一只会美牵球 我哼得起国 我灵魂失控 黑云在降落 我被它拖着走 如果我遇见你是一场悲剧 我可以让生\n",
+      " - 不分开吗 我叫你爸 你打我妈 这样对吗干嘛这样 何必让酒牵鼻子走 瞎 说也你在旧每 太的话说的忧剩 回说林\n",
+      "epoch 250, perplexity 1.336056, time 0.59 sec\n",
+      " - 分开 我爱想很样力 你作 失去开的玩笑 想通 却又再考倒我 说散 你想很久了吧? 我的认画败给黑色幽默 \n",
+      " - 不分开扫把的胖女巫 用拉丁文念咒语啦啦呜 她养的黑猫笑起来像哭 啦啦啦呜 刻多将痛 每一秒钟 你给没中 不\n"
+     ]
+    }
+   ],
+   "source": [
+    "train_and_predict_rnn(rnn, get_params, init_rnn_state, num_hiddens,\n",
+    "                      vocab_size, device, corpus_indices, idx_to_char,\n",
+    "                      char_to_idx, True, num_epochs, num_steps, lr,\n",
+    "                      clipping_theta, batch_size, pred_period, pred_len,\n",
+    "                      prefixes)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 43,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "epoch 50, perplexity 62.258360, time 0.60 sec\n",
+      " - 分开 我想要这 快使我有 我想一直 我想一空 我想一空 我想一空 我想一空 我想一空 我想一空 我想一空\n",
+      " - 不分开 我有你的可写 我有你的让写 我有你的 快使了双 我想一直 我想一空 我想一空 我想一空 我想一空 \n",
+      "epoch 100, perplexity 7.132725, time 0.59 sec\n",
+      " - 分开 一颗我 印地了人在江 还我在很医 我不要再想 我不 我不 我不要再想 我不 我不 我不要再想你 不\n",
+      " - 不分开觉你 我想要你想 我不要再想 我不 我不 我不要再想 我不 我不 我不要再想 我不 我不 我不要再想\n",
+      "epoch 150, perplexity 2.102736, time 0.65 sec\n",
+      " - 分开 问候我 谁地了的让 就颗莫衷身力 它所许秋 漫天黄沙凉过 塞北的客栈人多 牧草有没有 我马儿有些瘦\n",
+      " - 不分开觉 你已经离开我 不知不觉 我跟了这节奏 后知后觉 又过了一个秋 后生后觉 快使用双截棍 哼哼哈兮 \n",
+      "epoch 200, perplexity 1.317644, time 0.63 sec\n",
+      " - 分开 问候我 谁地我 陪你了那信堡我 甩开球我满腔的怒火 我想揍你已经很久 别想躲 说你眼睛看着我 别发\n",
+      " - 不分开觉 你已经离开我 不知不觉 我跟了这节奏 后知后觉 又过了一个秋 后知后觉 我该好好生活 我该好好生\n",
+      "epoch 250, perplexity 1.155166, time 0.70 sec\n",
+      " - 分开 问候我 谁地神枪在币 悲伤得的隐密 后录那这对里 藤蔓都靠我 你拿着球现投 又下会掩护我 选你这种\n",
+      " - 不分开觉 你已经离开我 不知不觉 我跟了这节奏 后知后觉 又过了一个秋 后知后觉 我该好好生活 我该好好生\n"
+     ]
+    }
+   ],
+   "source": [
+    "train_and_predict_rnn(rnn, get_params, init_rnn_state, num_hiddens,\n",
+    "                      vocab_size, device, corpus_indices, idx_to_char,\n",
+    "                      char_to_idx, False, num_epochs, num_steps, lr,\n",
+    "                      clipping_theta, batch_size, pred_period, pred_len,\n",
+    "                      prefixes)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "collapsed": true
+   },
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python [default]",
+   "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": "3.6.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}