TF2高阶操作

TF2高阶操作.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "TF2高阶操作",
      "provenance": [],
      "collapsed_sections": [],
      "toc_visible": true,
      "authorship_tag": "ABX9TyPHyuBfRr9HLFoUP2veMVy/",
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/gist/visualDust/358c4e9c7c654acc2a64a74c4688e44d/tf2.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "tr5--0KmJYes"
      },
      "source": [
        "# 合并与分割(Merge and split)  \n",
        "* tf.concat\n",
        "* tf.split\n",
        "* tf.stack\n",
        "* tf.unstack"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "waPKeLXlPaXH"
      },
      "source": [
        "# importing\n",
        "import os\n",
        "# changing env log level\n",
        "os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\n",
        "import tensorflow as tf\n",
        "from tensorflow.keras import layers\n",
        "from tensorflow.keras import optimizers"
      ],
      "execution_count": 2,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "XLzdji8OJvjZ"
      },
      "source": [
        "## concat\n",
        "concat用于张量的拼接操作。例如：一共有六个班级需要统计成绩。其中第一个人统计前四个班级的成绩，另一个人统计后两个班级的成绩。假设每个班有35人，每个人有八门科目的成绩，那么两个人获得的成绩单的shape应该分别是[4,35,8]和[2,35,8]，拼接后的成绩单的shape应该是[6,35,8]。"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "lVSJ_nAJJO5a",
        "outputId": "a9151573-0419-46b7-eeda-f3f1216e3f2b",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "a = tf.ones([4,35,8])\n",
        "b = tf.ones([2,35,8])\n",
        "# The operation of concat is seemingly two steps. 1. broadcast. 2. combine \n",
        "c = tf.concat([a,b],axis = 0)\n",
        "print(c.shape)"
      ],
      "execution_count": 3,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "(6, 35, 8)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "96WPotIHQCXh"
      },
      "source": [
        "另外一个类似的场景是两个人统计一个班级的成绩信息，该班级一共有35名学生，第一个人统计前32名学生的成绩，第二个人统计后3名学生的成绩，拼接后得到全班的总成绩单。"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "0IiGnOf4QVem",
        "outputId": "6d69b825-425d-4780-cfea-19f3fda26300",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "a = tf.ones([1,32,8])\n",
        "b = tf.ones([1,3,8])\n",
        "# The operation of concat is seemingly two steps. 1. broadcast. 2. combine \n",
        "c = tf.concat([a,b],axis = 1)\n",
        "print(c.shape)"
      ],
      "execution_count": 4,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "(1, 35, 8)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "DPFtTqW7Tqwi"
      },
      "source": [
        "当然也有这样的场景：现在有四个班级的人考了总共16门考试，其成绩分别记录在了两张表中，每张表记录了8门成绩。现在要将这些成绩放入同一张表："
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "pz67APchTrKb",
        "outputId": "9b3794d6-161c-4248-fb79-d96e123220b7",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "a = tf.ones([4,35,8])\n",
        "b = tf.ones([4,35,8])\n",
        "c = tf.concat([a,b],axis = -1)\n",
        "print(c.shape)"
      ],
      "execution_count": 5,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "(4, 35, 16)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "dc2WbmwXQbob"
      },
      "source": [
        "请注意这两个场景在运算时的区别是基于哪个维度进行拼接。第一种场景下对第0维度进行拼接，第二种场景下对第一维度进行拼接。  \n",
        "* concat的使用限制条件为：出了要拼接的维度的大小可以不等之外其它维度需要相等。"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "dg8ut_bQSbXn"
      },
      "source": [
        "## stack\n",
        "stack用于张量的堆叠操作。例如：现在有两个班级的成绩信息，张量结构为[class,student,scoer]。这两个班级分别属于两个学校，现在要将它们放入一张成绩表中，但是要能区分他们的学校。"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "iMNRV5ZGQruP",
        "outputId": "250b340f-3bc0-446f-d868-7b0a9eae00fc",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "a = tf.ones([4,35,8])\n",
        "b = tf.ones([4,35,8])\n",
        "# add a new dim and combine them\n",
        "c = tf.stack([a,b],axis = 0)\n",
        "print(c.shape)"
      ],
      "execution_count": 6,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "(2, 4, 35, 8)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "TZs7oJELUq4C"
      },
      "source": [
        "stack可选要扩展维度的位置，例如，我们希望将学校一列放在最后："
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Xqz1djfuUvOt",
        "outputId": "654f566b-00e0-4474-f7f4-8e7f29128ac9",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "a = tf.ones([4,35,8])\n",
        "b = tf.ones([4,35,8])\n",
        "# add a new dim and combine them\n",
        "c = tf.stack([a,b],axis = 3)\n",
        "print(c.shape)"
      ],
      "execution_count": 8,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "(4, 35, 8, 2)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "TsEcng9fVdq2"
      },
      "source": [
        "不过一般习惯上把更大的维度（学校）放在前面。  \n",
        "* stack的使用限制条件为shape相等。"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Btqr09g2Wtkr"
      },
      "source": [
        "## unstack\n",
        "对应stack，也有unstack。unstack可以在指定的axis上将tensor打散为该axis的size份"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "JebktaPDWtT_",
        "outputId": "ce7818e6-83aa-4c11-e8b5-825673311d1c",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "a = tf.ones([4,35,8])\n",
        "b = tf.ones([4,35,8])\n",
        "# add a new dim and combine them\n",
        "c = tf.stack([a,b],axis = 0)\n",
        "print(\"shape of the origin : \",c.shape)\n",
        "# unstack\n",
        "a_2,b_2 = tf.unstack(c,axis = 0)\n",
        "print(\"after unstack : a2:\",a_2.shape,\",b2:\",b_2.shape)"
      ],
      "execution_count": 9,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "shape of the origin :  (2, 4, 35, 8)\n",
            "after unstack : a2: (4, 35, 8) ,b2: (4, 35, 8)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "yPmuIestPS24"
      },
      "source": [
        "## split \n",
        "unstack的使用场景有限。split的功能更加强大。split大体上有两种用法：  \n",
        "* 第一种：num_or_size_splits是数字，例如\"num_or_size_splits=2\"的情况，split会将tensor再指定的axis上分成两半。  \n",
        "* 第二种：num_or_size_splits是一个list，例如\"num_or_size_splits=[1,2,3]\"的情况，split会将tensor在指定的axis上分为这个list的size份，在这里是3份，每份的相对大小分别是1、2、3。"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "WqGiWlWPPX37",
        "outputId": "c9869449-2d96-4096-a3fd-6c9ef0bafb97",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "a = tf.ones([4,35,8])\n",
        "b = tf.ones([4,35,8])\n",
        "# add a new dim and combine them\n",
        "c = tf.stack([a,b],axis = 0)\n",
        "print(\"shape of the origin : \",c.shape)\n",
        "# split into two part on axis 0\n",
        "res = tf.split(c,axis = 0,num_or_size_splits=2)\n",
        "print(\"after split into two part, len = \",len(res),\", shape = \",res[0].shape,\" and \",res[1].shape)"
      ],
      "execution_count": 10,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "shape of the origin :  (2, 4, 35, 8)\n",
            "after split into two part, len =  2 , shape =  (1, 4, 35, 8)  and  (1, 4, 35, 8)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "yRfteAaWRDl0",
        "outputId": "b51f805d-6d7b-4b51-dc4b-461e1c522832",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "a = tf.ones([4,35,8])\n",
        "b = tf.ones([4,35,8])\n",
        "# add a new dim and combine them\n",
        "c = tf.stack([a,b],axis = 0)\n",
        "print(\"shape of the origin : \",c.shape)\n",
        "# split into three part on axis 3, relative size = 2 ,2 ,4\n",
        "res = tf.split(c,axis = 3,num_or_size_splits=[2,2,4])"
      ],
      "execution_count": 11,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "shape of the origin :  (2, 4, 35, 8)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "iv7nxDvyaHYV"
      },
      "source": [
        "# 数据统计\n",
        "* tf.norm：张量范数（一范数、二范数、...、无穷范数）\n",
        "* tf.reduce_min：最小值\n",
        "* tf,reduce_max：最大值\n",
        "* tf.argmin：最小值位置\n",
        "* tf.argmax：最大值位置\n",
        "* tf.equal：张量比较\n",
        "* tf.unique：独特值"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "UBUKML45eX_2"
      },
      "source": [
        "## tf.norm  \n",
        "为了好理解，暂时只讨论向量的范数。向量的二范数的公式为：    \n",
        "$$^2\\sqrt{sum_{i=1}^{size}x_i^2}$$  \n",
        "向量的n范数的公式为：  \n",
        "$$^n\\sqrt{sum_{i=1}^{size}x_i^n}$$  \n",
        "可以理解为：范数是一个函数，是一个向量到数值的映射。向量之间无法比较大小，进行范数运算之后就能直接比较大小了。再换句话理解，这是一种特殊的\"欧氏距离(x)\"，可以比较向量到远点的距离(x)（我瞎理解的）。"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "__bO63gbe3rc",
        "outputId": "97648975-0ce5-4243-8cd1-0599d1fdf759",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "origin = tf.ones([2,2])\n",
        "# 二范数\n",
        "print(\"origin = \",origin,\"\\nafter norm: \",tf.norm(origin))\n",
        "# 验证一下二范数的运算方式和我们上面说的是否一致\n",
        "print(\"origin = \",origin,\"\\nafter square-square-aqrt : \",tf.sqrt(tf.reduce_sum(tf.square(origin))))\n",
        "print(\"They are the same\")"
      ],
      "execution_count": 12,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "origin =  tf.Tensor(\n",
            "[[1. 1.]\n",
            " [1. 1.]], shape=(2, 2), dtype=float32) \n",
            "after norm:  tf.Tensor(2.0, shape=(), dtype=float32)\n",
            "origin =  tf.Tensor(\n",
            "[[1. 1.]\n",
            " [1. 1.]], shape=(2, 2), dtype=float32) \n",
            "after square-square-aqrt :  tf.Tensor(2.0, shape=(), dtype=float32)\n",
            "They are the same\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "JcRv1-WphNh5",
        "outputId": "453bc006-bb13-46d3-bb68-5727347ac973",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "origin = tf.ones([4,28,28,3])\n",
        "# more complex example\n",
        "print(\"origin = \",origin.shape,\"\\nafter norm: \",tf.norm(origin)) \n",
        "# 验证一下二范数的运算方式和我们上面说的是否一致\n",
        "print(\"origin = \",origin.shape,\"\\nafter square-square-aqrt : \",tf.sqrt(tf.reduce_sum(tf.square(origin))))\n",
        "print(\"They are the same\")"
      ],
      "execution_count": 16,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "origin =  (4, 28, 28, 3) \n",
            "after norm:  tf.Tensor(96.99484, shape=(), dtype=float32)\n",
            "origin =  (4, 28, 28, 3) \n",
            "after square-square-aqrt :  tf.Tensor(96.99484, shape=(), dtype=float32)\n",
            "They are the same\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "9mo0TUQgCUMT"
      },
      "source": [
        "norm除了可以作用在整个张量上，也可以作用在某一个维度上。大概可以理解为对这个维度进行一次unstack然后再对unstack出来的每一个向量求norm。"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "AgpNVByuC419",
        "outputId": "19fc6cc8-73df-4d50-8e26-1958d6515298",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "origin = tf.ones([4,28,28,3])\n",
        "# norm working on specific axis\n",
        "print(\"origin = \",origin.shape,\"\\nafter norm on axis = 3 : \",tf.norm(origin,axis = 3))"
      ],
      "execution_count": 20,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "origin =  (4, 28, 28, 3) \n",
            "after norm on axis = 3 :  tf.Tensor(\n",
            "[[[1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  ...\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]]\n",
            "\n",
            " [[1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  ...\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]]\n",
            "\n",
            " [[1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  ...\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]]\n",
            "\n",
            " [[1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  ...\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]\n",
            "  [1.7320508 1.7320508 1.7320508 ... 1.7320508 1.7320508 1.7320508]]], shape=(4, 28, 28), dtype=float32)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "D10dcSG2EW46"
      },
      "source": [
        "除了默认的二范数外，norm也可以求n范数。方法是指定ord参数。例如："
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "NBYQNNIjEfyz",
        "outputId": "78703c1a-2381-4127-e181-8341118365de",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "origin = tf.ones([2,2])\n",
        "# 一范数\n",
        "print(\"ord = 1 : \",tf.norm(origin,ord=1))\n",
        "# 二范数\n",
        "print(\"ord = 2 : \",tf.norm(origin,ord=2))\n",
        "# 三范数\n",
        "print(\"ord = 3 : \",tf.norm(origin,ord=3))\n",
        "# 四范数\n",
        "print(\"ord = 4 : \",tf.norm(origin,ord=4))\n",
        "# 五范数\n",
        "print(\"ord = 5 : \",tf.norm(origin,ord=5))"
      ],
      "execution_count": 22,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "ord = 1 :  tf.Tensor(4.0, shape=(), dtype=float32)\n",
            "ord = 2 :  tf.Tensor(2.0, shape=(), dtype=float32)\n",
            "ord = 3 :  tf.Tensor(1.587401, shape=(), dtype=float32)\n",
            "ord = 4 :  tf.Tensor(1.4142135, shape=(), dtype=float32)\n",
            "ord = 5 :  tf.Tensor(1.319508, shape=(), dtype=float32)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Pl35JjaQGasK"
      },
      "source": [
        "## tf.reduce_min / max / mean / sum\n",
        "* tf.reduce_min\n",
        "* tf.reduce_max\n",
        "* tf.reduce_mean\n",
        "* tf.reduce_sum\n",
        "\n",
        "其实就是求最小值最大值平均值。名字里带着reduce表明，这个操作会有一个类似\"打平\"的过程。例如，当不指定axis参数时，一个[10,4]的tensor会被\"打平\"成一个[40]的\"list\"并求最大值、最小值....；再如，带有axis=2参数时，一个[10,4,10]的tensor会被\"降维\"变成一个元素为[10,4]的tensor的list，大小是10，然后对着十个元素进行最大、最小....运算。"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "imKwvLzTIext",
        "outputId": "f6e479b7-c103-4e1d-cc78-2d776777ebef",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "origin = tf.random.normal([4,10])\n",
        "print(\"origin = \",origin,\"\\nreduce_min = \",tf.reduce_min(origin),\"\\nreduce_max = \",tf.reduce_max(origin),\"\\nreduce_mean = \",tf.reduce_mean(origin))"
      ],
      "execution_count": 24,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "origin =  tf.Tensor(\n",
            "[[-1.4178391e+00  1.0799263e+00  1.8141992e+00 -2.9427743e-01\n",
            "   3.5776252e-01 -6.9446379e-01 -7.1207196e-01  9.6388352e-01\n",
            "  -2.1230397e+00  4.8318788e-01]\n",
            " [-4.1854006e-01 -2.2664030e-01 -9.8776561e-01  3.3819950e-01\n",
            "   2.4363371e-02 -3.2178679e+00 -2.8521428e-01 -5.3039378e-01\n",
            "  -1.0285269e+00 -1.2320877e+00]\n",
            " [ 6.0093373e-01  1.3320454e-02  9.5860285e-01  1.4495020e+00\n",
            "   5.1962131e-01  1.1331964e+00 -1.0149366e+00 -5.1126540e-02\n",
            "  -5.0443190e-01  3.9746460e-01]\n",
            " [-4.1444901e-01 -1.2171540e+00 -8.4814447e-01  1.4405949e+00\n",
            "   7.2787516e-04  1.2379333e+00  1.0925928e+00 -9.9176753e-01\n",
            "   3.8999468e-02  1.0164096e+00]], shape=(4, 10), dtype=float32) \n",
            "reduce_min =  tf.Tensor(-3.2178679, shape=(), dtype=float32) \n",
            "reduce_max =  tf.Tensor(1.8141992, shape=(), dtype=float32) \n",
            "reduce_mean =  tf.Tensor(-0.08123293, shape=(), dtype=float32)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "KePndTq3ML45",
        "outputId": "0ae30f7c-b02d-4b2b-e5cc-428eddb1c654",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "origin = tf.random.normal([4,10])\n",
        "print(\"origin = \",origin)\n",
        "print(\"\\nreduce_min on axis 1 = \",tf.reduce_min(origin,axis = 1))\n",
        "print(\"\\nreduce_max on axis 1 = \",tf.reduce_max(origin,axis = 1))\n",
        "print(\"\\nreduce_mean on axis 1 = \",tf.reduce_mean(origin,axis = 1))"
      ],
      "execution_count": 25,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "origin =  tf.Tensor(\n",
            "[[-3.1204236   0.67563623 -0.9232384   1.1589053   0.8515049  -0.47955766\n",
            "  -1.723766    0.12821583 -0.6078169  -0.07115268]\n",
            " [-0.03351626  0.5452725   0.4999855  -0.13481826  0.6798329   0.23792107\n",
            "  -0.6113948   1.3868407   0.24892737 -0.41333905]\n",
            " [-0.9676226  -0.3656622  -0.688232    1.721823    0.6695465  -0.44504106\n",
            "   0.90125936  0.5428907   1.4090685  -0.9626962 ]\n",
            " [-0.87203074  0.9285623   0.56897074 -1.4624474   1.8943952  -0.5554827\n",
            "  -0.8351434  -0.3565093  -1.5708245  -1.1640625 ]], shape=(4, 10), dtype=float32)\n",
            "\n",
            "reduce_min on axis 1 =  tf.Tensor([-3.1204236 -0.6113948 -0.9676226 -1.5708245], shape=(4,), dtype=float32)\n",
            "\n",
            "reduce_max on axis 1 =  tf.Tensor([1.1589053 1.3868407 1.721823  1.8943952], shape=(4,), dtype=float32)\n",
            "\n",
            "reduce_mean on axis 1 =  tf.Tensor([-0.4111693   0.24057117  0.18153341 -0.34245723], shape=(4,), dtype=float32)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "j_TVY0HPNPLb"
      },
      "source": [
        "## tf.argmax/argmin  \n",
        "* tf.argmax\n",
        "* tf.argmin\n",
        "\n",
        "用于求最小值和最大值的位置。当不指定axis参数时，默认再维度0上求每个维度下标下的最大、最小值的位置。"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "zR7Vtf61Nu7v",
        "outputId": "b6a7b6e0-c81c-429b-9a04-d407a94d3f7c",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "origin = tf.nn.relu(tf.random.normal([4,3])*100)\n",
        "print(\"origin = \",origin)\n",
        "# 维度0下求最大值位置\n",
        "print(\"argmax : \",tf.argmax(origin))"
      ],
      "execution_count": 33,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "origin =  tf.Tensor(\n",
            "[[ 0.        0.        0.      ]\n",
            " [ 0.        0.       54.485176]\n",
            " [86.468796 74.000046  0.      ]\n",
            " [ 0.       29.033602 69.07481 ]], shape=(4, 3), dtype=float32)\n",
            "argmax :  tf.Tensor([2 2 3], shape=(3,), dtype=int64)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Jl9brJX2PK38",
        "outputId": "0413e344-8e8c-4e30-87e5-d71e99cb8e92",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 0
        }
      },
      "source": [
        "origin = tf.nn.relu(tf.random.normal([4,3,2])*100)\n",
        "print(\"origin = \",origin)\n",
        "# 维度0下求最大值位置，这里对维度0展开会得到二位的tensor，所以得到的最值得位置也会是二维坐标\n",
        "print(\"argmax : \",tf.argmax(origin))"
      ],
      "execution_count": 34,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "origin =  tf.Tensor(\n",
            "[[[  0.        44.180485]\n",
            "  [ 60.554047  54.124874]\n",
            "  [ 11.455048  54.916447]]\n",
            "\n",
            " [[  0.        70.35009 ]\n",
            "  [147.33435  110.680046]\n",
            "  [  0.        59.37093 ]]\n",
            "\n",
            " [[  0.        20.160051]\n",
            "  [  0.        24.07408 ]\n",
            "  [  0.         0.      ]]\n",
            "\n",
            " [[ 84.53291    0.      ]\n",
            "  [131.28426  103.82523 ]\n",
            "  [192.91162   31.05115 ]]], shape=(4, 3, 2), dtype=float32)\n",
            "argmax :  tf.Tensor(\n",
            "[[3 1]\n",
            " [1 1]\n",
            " [3 1]], shape=(3, 2), dtype=int64)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "kKwO6iFZQb1c"
      },
      "source": [
        "## tf.equal\n",
        "用于比较"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "Xy3xNeBqXJ9J",
        "outputId": "ffe37aac-7c46-4964-effa-bf4f7eedc508",
        "colab": {
          "base_uri": "https://localhost:8080/"
        }
      },
      "source": [
        "a = tf.constant([1,2,3,4,5])\n",
        "b = tf.constant(range(5))\n",
        "print(\"a = \",a,\", b = \",b)\n",
        "result = tf.equal(a,b)\n",
        "print(\"Equal : \",result)\n",
        "cast_to_int = tf.reduce_sum(tf.cast(result, dtype=tf.int32))\n",
        "print(\"To int32 : \",cast_to_int)"
      ],
      "execution_count": 37,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "a =  tf.Tensor([1 2 3 4 5], shape=(5,), dtype=int32) , b =  tf.Tensor([0 1 2 3 4], shape=(5,), dtype=int32)\n",
            "Equal :  tf.Tensor([False False False False False], shape=(5,), dtype=bool)\n",
            "To int32 :  tf.Tensor(0, shape=(), dtype=int32)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "xPUymLEAZHzf"
      },
      "source": [
        "tf.equal在精确度计算过上似乎有点用。例如，当有一个测试数据集，你的模型跑出来的预测值和测试数据的y做一次equal，然后cast成一个数字，根据大小可以判断accuracy。（也就是相同的部分是准确预测的）"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "QWOFm0wDhTKm"
      },
      "source": [
        "## tf.unique  \n",
        "tf.unique能得到一个包含tensor中所有元素的“set”，并且得到另一个idx的tensor用户标注每一个元素在得到的set里的下标。例如："
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "hEjDBo8siJjZ",
        "outputId": "59c49ba0-3b02-49b6-d29c-b86a32a186dd",
        "colab": {
          "base_uri": "https://localhost:8080/"
        }
      },
      "source": [
        "origin = tf.constant([4,2,2,4,3])\n",
        "result, idx = tf.unique(origin) \n",
        "print(\"origin = \",origin,\"\\nunique : \",result,\", \\nidx : \",idx)"
      ],
      "execution_count": 5,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "origin =  tf.Tensor([4 2 2 4 3], shape=(5,), dtype=int32) \n",
            "unique :  tf.Tensor([4 2 3], shape=(3,), dtype=int32) , \n",
            "idx :  tf.Tensor([0 1 1 0 2], shape=(5,), dtype=int32)\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "We6BP_RfjP69"
      },
      "source": [
        "回忆一下tf2基本操作中的gather，我们可以通过得到的结果的得到的idx将它复原"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "nGRU51jujB8V",
        "outputId": "be9a8a66-4c1e-4b22-93b2-7765afa0f841",
        "colab": {
          "base_uri": "https://localhost:8080/"
        }
      },
      "source": [
        "origin = tf.constant([4,2,2,4,3])\n",
        "result, idx = tf.unique(origin) \n",
        "print(\"origin = \",origin,\"\\nunique : \",result,\", \\nidx : \",idx)\n",
        "print(\"using tf.gather : \",tf.gather(result,idx))"
      ],
      "execution_count": 6,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "origin =  tf.Tensor([4 2 2 4 3], shape=(5,), dtype=int32) \n",
            "unique :  tf.Tensor([4 2 3], shape=(3,), dtype=int32) , \n",
            "idx :  tf.Tensor([0 1 1 0 2], shape=(5,), dtype=int32)\n",
            "using tf.gather :  tf.Tensor([4 2 2 4 3], shape=(5,), dtype=int32)\n"
          ],
          "name": "stdout"
        }
      ]
    }
  ]
}