{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "06b0d2c9",
   "metadata": {},
   "source": [
    "# <center> SSD算法复现"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "30e46dc6",
   "metadata": {},
   "source": [
    "在上一课程vgg16中，我们复现了论文中的算法。在SSD算法中还不能直接拿vgg16作为特征提取器，我们需要修改一下模型。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8d1c9081",
   "metadata": {},
   "source": [
    "## 1.算法结构"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "03d9cb10",
   "metadata": {},
   "source": [
    "<img src='ssd.png' width=60% align=center>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9fe1a1cb",
   "metadata": {},
   "source": [
    "<img src='vgg_detail.png' width=70% align=center>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8df5108a",
   "metadata": {},
   "source": [
    "1.输入不再是224，在SSD算法中可以选择300或者512，称之为SSD300、SSD512。  \n",
    "2.第3个最大池化层使用的ceil而不是floor，其他不变。具体使用方法请参考nn.MaxPool2d。在conv3_3卷积后的特征图尺寸是$75\\times75$，使用ceil保证特征图的维度是$38\\times38$的偶数，否则是$37\\times37$的奇数，这样做的目的是方便处理。  \n",
    "3.把第5个最大池化层的核大小改成3，步长改成1。  \n",
    "4.第8层全连接层不要，对全连接层6、7改成卷积。  \n",
    "5.我们只需要关注conv4_3和conv7就可以了。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5fb20620",
   "metadata": {},
   "source": [
    "<img src=\"mod_vgg16.png\" width=80% align=center>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "9bef844f",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "from torchsummary import summary\n",
    "from torch.utils.data import DataLoader\n",
    "from torchvision.datasets import FashionMNIST\n",
    "from torchvision import models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "1bda593e",
   "metadata": {},
   "outputs": [],
   "source": [
    "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "ebfca277",
   "metadata": {},
   "outputs": [],
   "source": [
    "vgg_cfg = [(3, 3, 64), (3, 64, 64), 'M', (3, 64, 128), (3, 128, 128),\n",
    "       'M', (3, 128, 256), (3, 256, 256), (3, 256, 256), 'CM',\n",
    "       (3, 256, 512), (3, 512, 512), (3, 512, 512), 'M',(3, 512, 512),\n",
    "       (3, 512, 512), (3, 512, 512), '3M1', (3, 512, 1024, 6, 6), (1, 1024, 1024)\n",
    "      ]\n",
    "#M:2x2 2 floor mode-----CM: 2x2 2 ceil mode-----3M1: 3x3 1 floor mode"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "e103335f",
   "metadata": {},
   "outputs": [],
   "source": [
    "class BaseConv(nn.Module):\n",
    "    def __init__(self, kernel_size, in_channels, out_channels, padding=1, dilation=1, stride=1, act=True):\n",
    "        super(BaseConv, self).__init__()\n",
    "        self.conv = nn.Conv2d(kernel_size=kernel_size, in_channels=in_channels, out_channels=out_channels, padding=padding, dilation=dilation, stride=stride)\n",
    "        self.relu = nn.ReLU()\n",
    "        self.act = act\n",
    "    def forward(self, x):\n",
    "        if self.act:\n",
    "            return self.relu(self.conv(x))\n",
    "        else:\n",
    "            return self.conv(x)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "125c171c",
   "metadata": {},
   "outputs": [],
   "source": [
    "class VGG16(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(VGG16, self).__init__()\n",
    "        self.vgg_cfg = vgg_cfg\n",
    "        self.seq1 = self.net1()#获取conv4_3\n",
    "        self.seq2 = self.net2()#获取conv7\n",
    "        \n",
    "    def net1(self):\n",
    "        sequential = []\n",
    "        for c in self.vgg_cfg[: 13]:\n",
    "            if isinstance(c, tuple):\n",
    "                sequential.append(BaseConv(c[0], c[1], c[2], act=True))\n",
    "            else:\n",
    "                if c == 'M':\n",
    "                    sequential.append(nn.MaxPool2d(kernel_size=2, stride=2))\n",
    "                else:#CM\n",
    "                    sequential.append(nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True))\n",
    "        return nn.Sequential(*sequential)\n",
    "    \n",
    "    def net2(self):\n",
    "        sequential = []\n",
    "        for c in self.vgg_cfg[13: ]:\n",
    "            if isinstance(c, tuple):\n",
    "                if len(c) == 3:\n",
    "                    sequential.append(BaseConv(c[0], c[1], c[2], act=True))\n",
    "                else:\n",
    "                    sequential.append(BaseConv(c[0], c[1], c[2], c[3], c[4], act=True))\n",
    "            else:\n",
    "                if c == 'M':\n",
    "                    sequential.append(nn.MaxPool2d(kernel_size=2, stride=2))\n",
    "                else:#3M1\n",
    "                    sequential.append(nn.MaxPool2d(kernel_size=3, stride=1))\n",
    "        return nn.Sequential(*sequential)\n",
    "    \n",
    "    def forward(self, x):\n",
    "        conv4_3 = self.seq1(x)#图中的第4组中的第3个卷积\n",
    "        conv7 = self.seq2(conv4_3)#图中fc7替换的conv7\n",
    "        return conv4_3, conv7\n",
    "    \n",
    "    def load_pretrained_params(self):\n",
    "        current_dict = self.state_dict()#获取当前模型的名字和层\n",
    "        current_names = list(current_dict.keys())\n",
    "        pretrained_dict = models.vgg16(weights=models.VGG16_Weights.DEFAULT).state_dict()#pretrained已经改成weights，并且可以指定模型的版本\n",
    "        pretrained_names = list(pretrained_dict.keys())\n",
    "        #下面代码可以理解为迁移学习，把模型在其他数据集训练得到的参数赋给自己的模型\n",
    "        for i in range(len(pretrained_names[: -4])):#原VGG16中除掉后面两层全连接，只保留到fc6之前的，fc6，fc7还要转化为conv6，conv7\n",
    "            current_dict[current_names[i]] = pretrained_dict[pretrained_names[i]]\n",
    "        #fc6\n",
    "        fc6_weights = pretrained_dict['classifier.0.weight'].view(4096, 512, 7, 7)#(4096, 25088)--->(4096, 512, 7, 7)\n",
    "        fc6_bias = pretrained_dict['classifier.0.bias']#(4096)\n",
    "        #fc7\n",
    "        fc7_weights = pretrained_dict['classifier.3.weight'].view(4096, 4096, 1, 1)#(4096, 4096)--->(4096, 4096, 1, 1)\n",
    "        fc7_bias = pretrained_dict['classifier.3.bias']#(4096)\n",
    "        #subsample下采样conv6, conv7\n",
    "        #conv6\n",
    "        conv6_weights = self.decimate(fc6_weights, m=[4, None, 3, 3])#4096个通道，采样(0, 3, 7....)通道，7x7采样3x3(0, 2, 5),(4096,512,7,7）-->(1024, 512, 3, 3)\n",
    "        conv6_bias = self.decimate(fc6_bias, m=[4])#4096个通道采样4个通道(4096)-->(1024)\n",
    "        #conv7\n",
    "        conv7_weights = self.decimate(fc7_weights, m=[4, 4, None, None])#(4096, 4096, 1, 1)-->(1024, 1024, 1, 1)\n",
    "        conv7_bias = self.decimate(fc7_bias, m=[4])#4096个通道采样4个通道(4096)-->(1024)\n",
    "        #将采样后的权重复制给现在的网络conv6，conv7\n",
    "        current_dict['seq2.5.conv.weight'] = conv6_weights\n",
    "        current_dict['seq2.5.conv.bias'] = conv6_bias\n",
    "        current_dict['seq2.6.conv.weight'] = conv7_weights\n",
    "        current_dict[ 'seq2.6.conv.bias'] = conv7_bias\n",
    "        self.load_state_dict(current_dict)\n",
    "        print(\"pretrained params load finished!\")\n",
    "        \n",
    "    def decimate(self, tensor, m):\n",
    "        \"\"\"\n",
    "        Decimate a tensor by a factor 'm', i.e. downsample by keeping every 'm'th value.\n",
    "\n",
    "        This is used when we convert FC layers to equivalent Convolutional layers, BUT of a smaller size.\n",
    "\n",
    "        :param tensor: tensor to be decimated\n",
    "        :param m: list of decimation factors for each dimension of the tensor; None if not to be decimated along a dimension\n",
    "        :return: decimated tensor\n",
    "        \"\"\"\n",
    "        assert tensor.dim() == len(m)#判断维度是否正确\n",
    "        for d in range(tensor.dim()):\n",
    "            if m[d] is not None:\n",
    "                tensor = tensor.index_select(dim=d,\n",
    "                index=torch.arange(start=0, end=tensor.size(d), step=m[d]).long())#根据索引取出tensor\n",
    "        return tensor"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ea144c5a",
   "metadata": {},
   "source": [
    "vgg16每一层名字如下，一共32个，每一层包含权重和偏置，所以是16层。序号表示卷积位于整个模型的第几层，比如0过了是2，因为在0层卷积后面还有一个激活函数relu()，其他层下标都是这样计算。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "112b7c0c",
   "metadata": {},
   "source": [
    "官网的模型名字:['features.0.weight', 'features.0.bias', 'features.2.weight', 'features.2.bias', 'features.5.weight', 'features.5.bias', 'features.7.weight', 'features.7.bias', 'features.10.weight', 'features.10.bias', 'features.12.weight', 'features.12.bias', 'features.14.weight', 'features.14.bias', 'features.17.weight', 'features.17.bias', 'features.19.weight', 'features.19.bias', 'features.21.weight', 'features.21.bias', 'features.24.weight', 'features.24.bias', 'features.26.weight', 'features.26.bias', 'features.28.weight', 'features.28.bias', 'classifier.0.weight', 'classifier.0.bias', 'classifier.3.weight', 'classifier.3.bias', 'classifier.6.weight', 'classifier.6.bias']"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "11c2d3f8",
   "metadata": {},
   "source": [
    "conv6, conv7对应的层为:'seq2.5.conv.weight', 'seq2.5.conv.bias', 'seq2.6.conv.weight', 'seq2.6.conv.bias'"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "07564988",
   "metadata": {},
   "source": [
    "自己的模型名字:['seq1.0.conv.weight', 'seq1.0.conv.bias', 'seq1.1.conv.weight', 'seq1.1.conv.bias', 'seq1.3.conv.weight', 'seq1.3.conv.bias', 'seq1.4.conv.weight', 'seq1.4.conv.bias', 'seq1.6.conv.weight', 'seq1.6.conv.bias', 'seq1.7.conv.weight', 'seq1.7.conv.bias', 'seq1.8.conv.weight', 'seq1.8.conv.bias', 'seq1.10.conv.weight', 'seq1.10.conv.bias', 'seq1.11.conv.weight', 'seq1.11.conv.bias', 'seq1.12.conv.weight', 'seq1.12.conv.bias', 'seq2.1.conv.weight', 'seq2.1.conv.bias', 'seq2.2.conv.weight', 'seq2.2.conv.bias', 'seq2.3.conv.weight', 'seq2.3.conv.bias', 'seq2.5.conv.weight', 'seq2.5.conv.bias', 'seq2.6.conv.weight', 'seq2.6.conv.bias']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "883d01b4",
   "metadata": {},
   "outputs": [],
   "source": [
    "cls = VGG16()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "f225db7b",
   "metadata": {},
   "outputs": [],
   "source": [
    "x = torch.randn((1, 3, 300, 300))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "12fafd7a",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(torch.Size([1, 512, 38, 38]), torch.Size([1, 1024, 19, 19]))"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cls(x)[0].shape, cls(x)[1].shape"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b56f3146",
   "metadata": {},
   "source": [
    "除了VGG16作为基础模型以外，在最后一层还会增加辅助卷积神经网络。其目的是为了获取conv8_2，conv9_2， conv10_2，conv11_2的特征图。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0e95f3d0",
   "metadata": {},
   "source": [
    "<img src=\"auxiliary_conv.png\" width=80% align=center>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "9cf3dd8e",
   "metadata": {},
   "outputs": [],
   "source": [
    "auc_cfg = [(1, 1024, 256, 0, 1), (3, 256, 512, 1, 2), (1, 512, 128, 0, 1),\n",
    "          (3, 128, 256, 1, 2), (1, 256, 128, 0, 1), (3, 128, 256, 0, 1),\n",
    "          (1, 256, 128, 0, 1), (3, 128, 256, 0, 1)]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "81853b45",
   "metadata": {},
   "outputs": [],
   "source": [
    "class AuxiliaryConv(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(AuxiliaryConv, self).__init__()\n",
    "        self.auc_cfg = auc_cfg\n",
    "        self.fm = []\n",
    "        self.net()\n",
    "        self.init_param()#初始化参数\n",
    "        \n",
    "    def net(self):\n",
    "        sequential = []\n",
    "        for i in range(len(self.auc_cfg)):\n",
    "            c = self.auc_cfg[i]#获取参数\n",
    "            sequential.append(BaseConv(c[0], c[1], c[2], padding=c[3], stride=c[4], act=True))\n",
    "            if i % 2 == 1:#每2个为1组\n",
    "                self.fm.append(nn.Sequential(*sequential))\n",
    "                sequential = []\n",
    "    \n",
    "    def forward(self, conv7):\n",
    "        conv8_2 = self.fm[0](conv7)\n",
    "        conv9_2 = self.fm[1](conv8_2)\n",
    "        conv10_2 = self.fm[2](conv9_2)\n",
    "        conv11_2 = self.fm[3](conv10_2)\n",
    "        return conv8_2, conv9_2, conv10_2, conv11_2\n",
    "    \n",
    "    def init_param(self):\n",
    "        \"\"\"初始化参数\"\"\"\n",
    "        for c in self.children():\n",
    "            if isinstance(c, nn.Conv2d):\n",
    "                nn.init.xavier_normal_(c.weight)#初始化参数方法一般使用最后带_方法\n",
    "                nn.init.constant_(c.bias, 0.)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "2c185733",
   "metadata": {},
   "outputs": [],
   "source": [
    "auc = AuxiliaryConv()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "fc1cdf31",
   "metadata": {},
   "outputs": [],
   "source": [
    "x = torch.randn((1, 1024, 19, 19))#VGG16最后一层卷积特征图维度"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "287755f2",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(torch.Size([1, 512, 10, 10]),\n",
       " torch.Size([1, 256, 5, 5]),\n",
       " torch.Size([1, 256, 3, 3]),\n",
       " torch.Size([1, 256, 1, 1]))"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "auc(x)[0].shape, auc(x)[1].shape, auc(x)[2].shape, auc(x)[3].shape"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fac84e92",
   "metadata": {},
   "source": [
    "<img src=\"boxes.png\" width=70% align=center>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "82d4527c",
   "metadata": {},
   "outputs": [],
   "source": [
    "num_classes = 20#20个类别\n",
    "coords = 4 #4个坐标"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "d8880b00",
   "metadata": {},
   "outputs": [],
   "source": [
    "pred_cfg = [(3, 512, 4 * coords), (3, 1024, 6 * coords), (3, 512, 6 * coords),\n",
    "           (3, 256, 6 * coords), (3, 256, 4 * coords), (3, 256, 4 * coords),\n",
    "           (3, 512, 4 * num_classes), (3, 1024, 6 * num_classes), (3, 512, 6 * num_classes),\n",
    "           (3, 256, 6 * num_classes), (3, 256, 4 * num_classes), (3, 256, 4 * num_classes)]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "17aa68e1",
   "metadata": {},
   "outputs": [],
   "source": [
    "class Prediction(nn.Module):\n",
    "    def __init__(self, ):\n",
    "        super(Prediction, self).__init__()\n",
    "        self.num_classes = num_classes\n",
    "        self.pred_cfg = pred_cfg\n",
    "        self.fm = []\n",
    "        self.net()\n",
    "        self.init_param()\n",
    "        \n",
    "    def net(self):\n",
    "        for c in pred_cfg:\n",
    "            self.fm.append(BaseConv(c[0], c[1], c[2], act=False))\n",
    "            \n",
    "    def init_param(self):\n",
    "        \"\"\"初始化参数\"\"\"\n",
    "        for c in self.children():\n",
    "            if isinstance(c, nn.Conv2d):\n",
    "                nn.init.xavier_normal_(c.weight)#初始化参数方法一般使用最后带_方法\n",
    "                nn.init.constant_(c.bias, 0.)\n",
    "    \n",
    "    def forward(self, conv4_3, conv7, conv8_2, conv9_2, conv10_2, conv11_2):\n",
    "        batch_size = conv4_3.size(0)#获取批次\n",
    "        #conv4_3坐标处理\n",
    "        loc_conv4_3 = self.fm[0](conv4_3)#(-1, 16, 38, 38)\n",
    "        loc_conv4_3 = loc_conv4_3.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 38, 38, 16)\n",
    "        loc_conv4_3 = loc_conv4_3.view(batch_size, -1, 4)#(N, 5776, 4)\n",
    "        #conv7坐标处理\n",
    "        loc_conv7 = self.fm[1](conv7)#(-1, 24, 19, 19)\n",
    "        loc_conv7 = loc_conv7.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 19, 19, 24)\n",
    "        loc_conv7 = loc_conv7.view(batch_size, -1, 4)#(N, 2166, 4)\n",
    "        #conv8_2坐标处理\n",
    "        loc_conv8_2 = self.fm[2](conv8_2)#(-1, 24, 10, 10)\n",
    "        loc_conv8_2 = loc_conv8_2.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 10, 10, 24)\n",
    "        loc_conv8_2 = loc_conv8_2.view(batch_size, -1, 4)#(N, 600, 4)\n",
    "        #conv9_2坐标处理\n",
    "        loc_conv9_2 = self.fm[3](conv9_2)#(-1, 24, 5, 5)\n",
    "        loc_conv9_2 = loc_conv9_2.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 5, 5, 24)\n",
    "        loc_conv9_2 = loc_conv9_2.view(batch_size, -1, 4)#(N, 150, 4)\n",
    "        #conv10_2坐标处理\n",
    "        loc_conv10_2 = self.fm[4](conv10_2)#(-1, 16, 3, 3)\n",
    "        loc_conv10_2 = loc_conv10_2.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 3, 3, 16)\n",
    "        loc_conv10_2 = loc_conv10_2.view(batch_size, -1, 4)#(N, 36, 4)\n",
    "        #conv11_2坐标处理\n",
    "        loc_conv11_2 = self.fm[5](conv11_2)#(-1, 16, 1, 1)\n",
    "        loc_conv11_2 = loc_conv11_2.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 1, 1, 16)\n",
    "        loc_conv11_2 = loc_conv11_2.view(batch_size, -1, 4)#(N, 4, 4)\n",
    "        locations = torch.cat([loc_conv4_3, loc_conv7, loc_conv8_2, loc_conv9_2, loc_conv10_2, loc_conv11_2], dim=1)\n",
    "        \n",
    "        \n",
    "        #conv4_3类别处理\n",
    "        cls_conv4_3 = self.fm[6](conv4_3)#(-1, 4 * num_classes, 38, 38)\n",
    "        cls_conv4_3 = cls_conv4_3.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 38, 38, 4 * num_classes)\n",
    "        cls_conv4_3 = cls_conv4_3.view(batch_size, -1, num_classes)#(N, 5776, num_classes)\n",
    "        #conv7类别处理\n",
    "        cls_conv7 = self.fm[7](conv7)#(-1, 6 * num_classes, 19, 19)\n",
    "        cls_conv7 = cls_conv7.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 19, 19, 6 * num_classes)\n",
    "        cls_conv7 = cls_conv7.view(batch_size, -1, num_classes)#(N, 2166, num_classes)\n",
    "        #conv8_2类别处理\n",
    "        cls_conv8_2 = self.fm[8](conv8_2)#(-1, 6 * num_classes, 10, 10)\n",
    "        cls_conv8_2 = cls_conv8_2.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 10, 10, 6 * num_classes)\n",
    "        cls_conv8_2 = cls_conv8_2.view(batch_size, -1, num_classes)#(N, 600, num_classes)\n",
    "        #conv9_2类别处理\n",
    "        cls_conv9_2 = self.fm[9](conv9_2)#(-1, 6 * num_classes, 5, 5)\n",
    "        cls_conv9_2 = cls_conv9_2.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 5, 5, 6 * num_classes)\n",
    "        cls_conv9_2 = cls_conv9_2.view(batch_size, -1, num_classes)#(N, 150, num_classes)\n",
    "        #conv10_2类别处理\n",
    "        cls_conv10_2 = self.fm[10](conv10_2)#(-1, 4 * num_classes, 3, 3)\n",
    "        cls_conv10_2 = cls_conv10_2.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 3, 3, 4 * num_classes)\n",
    "        cls_conv10_2 = cls_conv10_2.view(batch_size, -1, num_classes)#(N, 36, num_classes)\n",
    "        #conv11_2类别处理\n",
    "        cls_conv11_2 = self.fm[11](conv11_2)#(-1, 4 * num_classes, 1, 1)\n",
    "        cls_conv11_2 = cls_conv11_2.permute(0, 2, 3, 1).contiguous()#contiguous使得内存中数据改变(-1, 1, 1, 4 * num_classes)\n",
    "        cls_conv11_2 = cls_conv11_2.view(batch_size, -1, num_classes)#(N, 4, num_classes)\n",
    "        locations = torch.cat([loc_conv4_3, loc_conv7, loc_conv8_2, loc_conv9_2, loc_conv10_2, loc_conv11_2], dim=1)\n",
    "        classes = torch.cat([cls_conv4_3, cls_conv7, cls_conv8_2, cls_conv9_2, cls_conv10_2, cls_conv11_2], dim=1)\n",
    "        return locations, classes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "dbe28eff",
   "metadata": {},
   "outputs": [],
   "source": [
    "p = Prediction()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "5ecde33e",
   "metadata": {},
   "outputs": [],
   "source": [
    "conv4_3 = torch.randn((1, 512, 38, 38))\n",
    "conv7 = torch.randn((1, 1024, 19, 19))\n",
    "conv8_2 = torch.randn((1, 512, 10, 10))\n",
    "conv9_2 = torch.randn((1, 256, 5, 5))\n",
    "conv10_2 = torch.randn((1, 256, 3, 3))\n",
    "conv11_2 = torch.randn((1, 256, 1, 1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "579114f2",
   "metadata": {},
   "outputs": [],
   "source": [
    "out = p(conv4_3, conv7, conv8_2, conv9_2, conv10_2, conv11_2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "012ac090",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(torch.Size([1, 8732, 4]), torch.Size([1, 8732, 20]))"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "out[0].shape, out[1].shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "86d16dc4",
   "metadata": {},
   "outputs": [],
   "source": [
    "x = nn.Parameter(torch.FloatTensor(1,100))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "id": "5f306349",
   "metadata": {},
   "outputs": [],
   "source": [
    "fm_dims = {'conv4_3': 38,\n",
    "          'conv7': 19,\n",
    "          'conv8_2': 10,\n",
    "          'conv9_2': 5,\n",
    "          'conv10_2': 3,\n",
    "          'conv11_2': 1}\n",
    "boxes_scale = {'conv4_3': 0.1,\n",
    "               'conv7': 0.2, \n",
    "               'conv8_2': 0.375,\n",
    "               'conv9_2': 0.55,\n",
    "               'conv10_2': 0.725,\n",
    "               'conv11_2': 0.9}\n",
    "aspect_ratios = {'conv4_3': [1., 2., 0.5],\n",
    "                 'conv7': [1., 2., 3., 0.5, 0.333],\n",
    "                 'conv8_2': [1., 2., 3., 0.5, 0.333],\n",
    "                 'conv9_2': [1., 2., 3., 0.5, 0.333],\n",
    "                 'conv10_2': [1., 2., 0.5],\n",
    "                 'conv11_2': [1., 2., 0.5]}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "id": "ba54e968",
   "metadata": {},
   "outputs": [
    {
     "ename": "SyntaxError",
     "evalue": "invalid character in identifier (3965864224.py, line 12)",
     "output_type": "error",
     "traceback": [
      "\u001b[1;36m  File \u001b[1;32m\"C:\\Users\\Stark-lin\\AppData\\Local\\Temp\\ipykernel_6056\\3965864224.py\"\u001b[1;36m, line \u001b[1;32m12\u001b[0m\n\u001b[1;33m    def forward(self， image):\u001b[0m\n\u001b[1;37m                    ^\u001b[0m\n\u001b[1;31mSyntaxError\u001b[0m\u001b[1;31m:\u001b[0m invalid character in identifier\n"
     ]
    }
   ],
   "source": [
    "class SSD(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(SSD, self).__init__()\n",
    "        self.fm_dims = fm_dims\n",
    "        self.boxes_scale = boxes_scale\n",
    "        self.aspect_ratios = aspect_ratios\n",
    "        self.base = VGG16()\n",
    "        self.auc = AuxiliaryConv()\n",
    "        self.prediction = Prediction()\n",
    "        self.rescale = nn.Parameter(torch.FloatTensor(1, 512, 1, 1))#对conv4_3进行缩放，低特征提取器的规模更大使用L2 norm归一化和其他层保持一个scale\n",
    "        nn.init.constant_(self.rescale, 20)#全是20，后面通过反向传播训练，nn.Parameter是可以更新的\n",
    "        \n",
    "    def forward(self， image):\n",
    "        conv4_3, conv7 = self.base(image)#(N, 512, 38, 38),(N, 1024, 19, 19)\n",
    "        norm = conv4_3.pow(2).sum(dim=1, keepdim=True).sqrt()#(N, 1, 38, 38)\n",
    "        conv4_3 = conv4_3 / norm #(N, 512, 38, 38)\n",
    "        conv4_3 = conv4_3 * self.rescale\n",
    "        \n",
    "        conv8_2, conv_9_2, conv10_2, conv11_2 = self.auc(conv_7)#(N, 512, 10, 10),(N, 256, 5, 5),(N, 256, 3, 3), (N, 256, 1, 1)\n",
    "        locations, classes = self.prediction(conv4_3, conv7, conv8_2, conv9_2, conv10_2, conv11_2)\n",
    "        return locations, classes\n",
    "    \n",
    "    def get_prior_boxes(self):\n",
    "        "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0b5cc276",
   "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": "3.7.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}