{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "3e1b8170",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-03-15T14:53:42.582632Z",
     "start_time": "2023-03-15T14:53:40.053444Z"
    }
   },
   "outputs": [],
   "source": [
    "import torch"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "895b4a69",
   "metadata": {},
   "source": [
    "## 基本介绍"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "55041716",
   "metadata": {},
   "source": [
    "- 复杂的接口，一些抽象的接口的深入理解\n",
    "    - 理解复杂的代码；\n",
    "    - 有助于构建复杂的神经网络，且清晰，不出错；\n",
    "    \n",
    "- 基于 index 在 原始 tensor 上选择（select）\n",
    "    - 行选多个（行选1个是一种特例），提供的是列值(`dim=1`)；\n",
    "        - 多个时，是几个必须一致；\n",
    "    - 列选多个（列选1个是一种特例），提供的是行值（`dim=0`）；\n",
    "\n",
    "- 要求，\n",
    "    - input tensor 与 index 的 dim 相同\n",
    "    - index.shape < input.shape\n",
    "\n",
    "- 应用场景\n",
    "    - batch sample 多分类问题的分类输出，获得各个true label 上的 score(logits)/probs（行选1个）\n",
    "        - batch sample: n*c\n",
    "        - index(true label): n*1\n",
    "    - dqn：batch state 的 batch action 的输出，回归问题；\n",
    "        ```\n",
    "        state_action_values = policy_net(state_batch).gather(1, action_batch)\n",
    "        ```\n",
    "        - state_batch: n*d\n",
    "        - action_batch: n*1\n",
    "        - state_action_values: n*1"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "78106c0f",
   "metadata": {},
   "source": [
    "## 官网示例"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "ce2df535",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-03-15T15:28:55.429980Z",
     "start_time": "2023-03-15T15:28:55.425003Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[1, 2],\n",
       "        [3, 4]])"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "t = torch.tensor([[1, 2], \n",
    "                  [3, 4]])\n",
    "t"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "1d63e435",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-03-15T15:29:32.016891Z",
     "start_time": "2023-03-15T15:29:32.011386Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[1, 1],\n",
       "        [4, 3]])"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "torch.gather(t, 1, torch.tensor([[0, 0], \n",
    "                                 [1, 0]]))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "89f41a09",
   "metadata": {},
   "source": [
    "[0, 0]\\\n",
    "[1, 0] => \n",
    "> [(0, 0), (0, 0)]\\\n",
    "> [(1, 1), (1, 0)]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "913c87ac",
   "metadata": {},
   "source": [
    "## 复杂例子"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "0fac917e",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-03-15T15:30:24.244197Z",
     "start_time": "2023-03-15T15:30:24.238490Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[1, 2, 3],\n",
       "        [4, 5, 6],\n",
       "        [7, 8, 9]])"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "t = torch.arange(1, 10).view(3, 3)\n",
    "t"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5d90c9d6",
   "metadata": {},
   "source": [
    "- 如果想选择 (1,3), (5, 6)\n",
    "    - 0, 2\n",
    "    - 1, 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "c422f32e",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-03-15T15:31:16.503161Z",
     "start_time": "2023-03-15T15:31:16.497969Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[1, 3],\n",
       "        [5, 6]])"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "t.gather(1, torch.tensor(((0, 2), (1, 2))))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fdfd1c88",
   "metadata": {},
   "source": [
    "- 如果想选择 (1, 4), (2, 8)，提供的是行值\n",
    "    - (0, 1), \n",
    "    - (0, 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "b4199432",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2023-03-15T15:32:01.607615Z",
     "start_time": "2023-03-15T15:32:01.602219Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[1, 2],\n",
       "        [4, 8]])"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "t.gather(0, torch.tensor(((0, 0), (1, 2))))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1a02c334",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.9.13"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {},
   "toc_section_display": true,
   "toc_window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}