09-theano/09.02-theano-basics.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Theano 基础"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "首先导入 `theano` 及其 `tensor` 子模块（`tensor`，张量）："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using gpu device 1: Tesla K10.G2.8GB (CNMeM is disabled)\n"
     ]
    }
   ],
   "source": [
    "import theano\n",
    "\n",
    "# 一般都把 `tensor` 子模块导入并命名为 T\n",
    "import theano.tensor as T"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "`tensor` 模块包含很多我们常用的数学操作，所以为了方便，将其命名为 T。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 符号计算"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "`theano` 中，所有的算法都是用符号计算的，所以某种程度上，用 `theano` 写算法更像是写数学（之前在[04.06 积分](../04. scipy/04.06 integration in python.ipynb)一节中接触过用 `sympy` 定义的符号变量）。\n",
    "\n",
    "用 `T.scalar` 来定义一个符号标量："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "foo = T.scalar('x')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "x\n"
     ]
    }
   ],
   "source": [
    "print foo"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "支持符号计算："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Elemwise{pow,no_inplace}.0\n"
     ]
    }
   ],
   "source": [
    "bar = foo ** 2\n",
    "\n",
    "print bar"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "这里定义 `foo` 是 $x$，`bar` 就是变量 $x^2$，但显示出来的却是看不懂的东西。\n",
    "\n",
    "为了更好的显示 `bar`，我们使用 `theano.pp()` 函数（`pretty print`）来显示："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(x ** TensorConstant{2})\n"
     ]
    }
   ],
   "source": [
    "print theano.pp(bar)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "查看类型："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "<class 'theano.tensor.var.TensorVariable'>\n",
      "TensorType(float32, scalar)\n"
     ]
    }
   ],
   "source": [
    "print type(foo)\n",
    "print foo.type"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## theano 函数"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "有了符号变量，自然可以用符号变量来定义函数，`theano.function()` 函数用来生成符号函数：\n",
    "\n",
    "    theano.function(input, output)\n",
    "\n",
    "其中 `input` 对应的是作为参数的符号变量组成的列表，`output` 对应的是输出，输出可以是一个，也可以是多个符号变量组成的列表。\n",
    "\n",
    "例如，我们用刚才生成的 `foo` 和 `bar` 来定义函数："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "square = theano.function([foo], bar)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "使用 `square` 函数："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "9.0\n"
     ]
    }
   ],
   "source": [
    "print square(3)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "也可以使用 `bar` 的 `eval` 方法，将 `x` 替换为想要的值，`eval` 接受一个字典作为参数，键值对表示符号变量及其对应的值："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "9.0\n"
     ]
    }
   ],
   "source": [
    "print bar.eval({foo: 3})"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## theano.tensor"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "除了 `T.scalar()` 标量之外，`Theano` 中还有很多符号变量类型，这些都包含在 `tensor`（张量）子模块中，而且 `tensor` 中也有很多函数对它们进行操作。\n",
    "\n",
    "- `T.scalar(name=None, dtype=config.floatX)` \n",
    "    - 标量，shape - ()\n",
    "- `T.vector(name=None, dtype=config.floatX)` \n",
    "    - 向量，shape - (?,)\n",
    "- `T.matrix(name=None, dtype=config.floatX)` \n",
    "    - 矩阵，shape - (?,?)\n",
    "- `T.row(name=None, dtype=config.floatX)` \n",
    "    - 行向量，shape - (1,?)\n",
    "- `T.col(name=None, dtype=config.floatX)` \n",
    "    - 列向量，shape - (?,1)\n",
    "- `T.tensor3(name=None, dtype=config.floatX)`\n",
    "    - 3 维张量，shape - (?,?,?)\n",
    "- `T.tensor4(name=None, dtype=config.floatX)`\n",
    "    - 4 维张量，shape - (?,?,?,?)\n",
    "\n",
    "`shape` 中为 1 的维度支持 `broadcast` 机制。\n",
    "\n",
    "除了直接指定符号变量的类型（默认 `floatX`），还可以直接在每类前面加上一个字母来定义不同的类型：\n",
    "\n",
    "- `b` int8\n",
    "- `w` int16\n",
    "- `i` int32\n",
    "- `l` int64\n",
    "- `d` float64\n",
    "- `f` float32\n",
    "- `c` complex64\n",
    "- `z` complex128\n",
    "\n",
    "例如 `T.dvector()` 表示的就是一个 `float64` 型的向量。\n",
    "\n",
    "除此之外，还可以用它们的复数形式一次定义多个符号变量：\n",
    "\n",
    "    x,y,z = T.vectors('x','y','z')\n",
    "    x,y,z = T.vectors(3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "A = T.matrix('A')\n",
    "x = T.vector('x')\n",
    "b = T.vector('b')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "`T.dot()` 表示矩阵乘法：\n",
    "$$y = Ax+b$$"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "y = T.dot(A, x) + b"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "`T.sum()` 表示进行求和：\n",
    "$$z = \\sum_{i,j} A_{ij}^2$$"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "z = T.sum(A**2)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "来定义一个线性函数，以 $A,x,b$ 为参数，以 $y,z$ 为输出： "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "linear_mix = theano.function([A, x, b],\n",
    "                             [y, z])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "使用这个函数：\n",
    "\n",
    "$$\n",
    "A = \\begin{bmatrix}\n",
    "1 & 2 & 3 \\\\\n",
    "4 & 5 & 6\n",
    "\\end{bmatrix}, \n",
    "x = \\begin{bmatrix}\n",
    "1 \\\\ 2 \\\\ 3\n",
    "\\end{bmatrix},\n",
    "b = \\begin{bmatrix}\n",
    "4 \\\\ 5\n",
    "\\end{bmatrix}\n",
    "$$"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[array([ 18.,  37.], dtype=float32), array(91.0, dtype=float32)]\n"
     ]
    }
   ],
   "source": [
    "print linear_mix(np.array([[1, 2, 3],\n",
    "                           [4, 5, 6]], dtype=theano.config.floatX),    #A\n",
    "                 np.array([1, 2, 3], dtype=theano.config.floatX),      #x\n",
    "                 np.array([4, 5], dtype=theano.config.floatX))         #b"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "这里 `dtype=theano.config.floatX` 是为了与 `theano` 设置的浮点数精度保持一致，默认是 `float64`，但是在 `GPU` 上一般使用 `float32` 会更高效一些。\n",
    "\n",
    "我们还可以像定义普通函数一样，给 `theano` 函数提供默认值，需要使用 `theano.Param` 类："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "linear_mix_default = theano.function([A, x, theano.Param(b, default=np.zeros(2, dtype=theano.config.floatX))],\n",
    "                                     [y, z])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "计算默认参数下的结果："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[array([ 14.,  32.], dtype=float32), array(91.0, dtype=float32)]\n"
     ]
    }
   ],
   "source": [
    "print linear_mix_default(np.array([[1, 2, 3],\n",
    "                           [4, 5, 6]], dtype=theano.config.floatX),    #A\n",
    "                 np.array([1, 2, 3], dtype=theano.config.floatX))      #x"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "计算刚才的结果："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[array([ 18.,  37.], dtype=float32), array(91.0, dtype=float32)]\n"
     ]
    }
   ],
   "source": [
    "print linear_mix_default(np.array([[1, 2, 3],\n",
    "                                   [4, 5, 6]], dtype=theano.config.floatX),    #A\n",
    "                         np.array([1, 2, 3], dtype=theano.config.floatX),      #x\n",
    "                         np.array([4, 5], dtype=theano.config.floatX))         #b"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 共享的变量"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "`Theano` 中可以定义共享的变量，它们可以在多个函数中被共享，共享变量类似于普通函数定义时候使用的全局变量，同时加上了 `global` 的属性以便在函数中修改这个全局变量的值。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CudaNdarrayType(float32, matrix)\n"
     ]
    }
   ],
   "source": [
    "shared_var = theano.shared(np.array([[1.0, 2.0], [3.0, 4.0]], dtype=theano.config.floatX))\n",
    "\n",
    "print shared_var.type"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "可以通过 `set_value` 方法改变它的值："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "shared_var.set_value(np.array([[3.0, 4], [2, 1]], dtype=theano.config.floatX))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "通过 `get_value()` 方法返回它的值："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[ 3.  4.]\n",
      " [ 2.  1.]]\n"
     ]
    }
   ],
   "source": [
    "print shared_var.get_value()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "共享变量进行运算："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[  9.  16.]\n",
      " [  4.   1.]]\n"
     ]
    }
   ],
   "source": [
    "shared_square = shared_var ** 2\n",
    "\n",
    "f = theano.function([], shared_square)\n",
    "\n",
    "print f()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "这里函数不需要参数，因为共享变量隐式地被认为是一个参数。\n",
    "\n",
    "得到的结果会随这个共享变量的变化而变化："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[  1.   4.]\n",
      " [  9.  16.]]\n"
     ]
    }
   ],
   "source": [
    "shared_var.set_value(np.array([[1.0, 2], [3, 4]], dtype=theano.config.floatX))\n",
    "\n",
    "print f()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "一个共享变量的值可以用 `updates` 关键词在 `theano` 函数中被更新："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "subtract = T.matrix('subtract')\n",
    "\n",
    "f_update = theano.function([subtract], shared_var, updates={shared_var: shared_var - subtract})"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "这个函数先返回当前的值，然后将当前值更新为原来的值减去参数："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "before update:\n",
      "[[ 1.  2.]\n",
      " [ 3.  4.]]\n",
      "the return value:\n",
      "<CudaNdarray object at 0x7f7f3c16a6f0>\n",
      "after update:\n",
      "[[ 0.  1.]\n",
      " [ 2.  3.]]\n"
     ]
    }
   ],
   "source": [
    "print 'before update:'\n",
    "print shared_var.get_value()\n",
    "\n",
    "print 'the return value:'\n",
    "print f_update(np.array([[1.0, 1], [1, 1]], dtype=theano.config.floatX))\n",
    "\n",
    "print 'after update:'\n",
    "print shared_var.get_value()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 导数"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "`Theano` 的一大好处在于它对符号变量计算导数的能力。\n",
    "\n",
    "我们用 `T.grad()` 来计算导数，之前我们定义了 `foo` 和 `bar` （分别是 $x$ 和 $x^2$）,我们来计算 `bar` 关于 `foo` 的导数（应该是 $2x$）："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "20.0\n"
     ]
    }
   ],
   "source": [
    "bar_grad = T.grad(bar, foo)  # 表示 bar (x^2) 关于 foo (x) 的导数\n",
    "\n",
    "print bar_grad.eval({foo: 10})"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "再如，对之前的 $y = Ax + b$ 求 $y$ 关于 $x$ 的雅可比矩阵（应当是 $A$）："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[ 9.  8.  7.]\n",
      " [ 4.  5.  6.]]\n"
     ]
    }
   ],
   "source": [
    "y_J = theano.gradient.jacobian(y, x)\n",
    "\n",
    "print y_J.eval({A: np.array([[9.0, 8, 7], [4, 5, 6]], dtype=theano.config.floatX), #A\n",
    "                x: np.array([1.0, 2, 3], dtype=theano.config.floatX),              #x\n",
    "                b: np.array([4.0, 5], dtype=theano.config.floatX)})                #b"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "`theano.gradient.jacobian` 用来计算雅可比矩阵，而 `theano.gradient.hessian` 可以用来计算 `Hessian` 矩阵。"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## `R-op` 和 `L-op`"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Rop 用来计算 $\\frac{\\partial f}{\\partial x}v$，Lop 用来计算 $v\\frac{\\partial f}{\\partial x}$：\n",
    "\n",
    "一个是雅可比矩阵与列向量的乘积，另一个是行向量与雅可比矩阵的乘积。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[ 2.  2.]\n"
     ]
    }
   ],
   "source": [
    "W = T.dmatrix('W')\n",
    "V = T.dmatrix('V')\n",
    "x = T.dvector('x')\n",
    "y = T.dot(x, W)\n",
    "JV = T.Rop(y, W, V)\n",
    "f = theano.function([W, V, x], JV)\n",
    "\n",
    "print f([[1, 1], [1, 1]], [[2, 2], [2, 2]], [0,1])"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}