Capsule_Pytorch.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "Capsule Pytorch.ipynb",
      "version": "0.3.2",
      "provenance": [],
      "collapsed_sections": []
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.6.8"
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "code",
      "metadata": {
        "id": "77gWhdAk8Z0_",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "import torch\n",
        "import torch.nn as nn\n",
        "import torch.nn.functional as F\n",
        "import torch.optim as optim\n",
        "from torchvision import datasets, transforms\n",
        "from torch.nn.init import xavier_uniform_ as init\n",
        "from torch.autograd import Variable\n",
        "import numpy as np"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "4RubMgSs8Z1J",
        "colab_type": "text"
      },
      "source": [
        "# 加载mnist数据集"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "p6ntKBnW8Z1J",
        "colab_type": "code",
        "outputId": "4283d776-6b4f-4055-f63a-d682cc2bbef6",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 35
        }
      },
      "source": [
        "BATCH_SIZE=128\n",
        "EPOCHS=20\n",
        "DEVICE = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
        "DEVICE"
      ],
      "execution_count": 2,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "device(type='cuda')"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 2
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "AS0ldjbc8Z1M",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "train_loader = torch.utils.data.DataLoader(\n",
        "        datasets.MNIST('data', train=True, download=True, \n",
        "                       transform=transforms.Compose([\n",
        "                           transforms.ToTensor(),\n",
        "                           transforms.Normalize((0.1307,), (0.3081,))\n",
        "                       ])),\n",
        "        batch_size=BATCH_SIZE, shuffle=True)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "PwGmQdt_8Z1O",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "test_loader = torch.utils.data.DataLoader(\n",
        "        datasets.MNIST('data', train=False, transform=transforms.Compose([\n",
        "                           transforms.ToTensor(),\n",
        "                           transforms.Normalize((0.1307,), (0.3081,))\n",
        "                       ])),\n",
        "        batch_size=BATCH_SIZE, shuffle=True)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "LT-93_sze85s",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "test_iter = iter(test_loader)\n",
        "imgs,labels = next(test_iter)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "htTReQjgiimg",
        "colab_type": "code",
        "outputId": "d16eb9dc-470e-4354-f9b6-96d14854dc75",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 35
        }
      },
      "source": [
        "print(imgs.shape,labels.shape)"
      ],
      "execution_count": 6,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "torch.Size([128, 1, 28, 28]) torch.Size([128])\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "LFqtb-Ajf1ug",
        "colab_type": "code",
        "outputId": "b593187c-befb-4979-f773-1a3c36a1dcdb",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 128
        }
      },
      "source": [
        "x_test = torch.tensor(imgs[:500],device=DEVICE,dtype=torch.float)\n",
        "# x_test = x_test.view(x_test.size(0),1, x_test.size(1),x_test.size(2))\n",
        "y_test = torch.tensor(labels[:500],device=DEVICE,dtype=torch.long)\n",
        "\n",
        "\n",
        "print(x_test.shape, y_test.shape)"
      ],
      "execution_count": 7,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "/usr/local/lib/python3.6/dist-packages/ipykernel_launcher.py:1: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
            "  \"\"\"Entry point for launching an IPython kernel.\n"
          ],
          "name": "stderr"
        },
        {
          "output_type": "stream",
          "text": [
            "torch.Size([128, 1, 28, 28]) torch.Size([128])\n"
          ],
          "name": "stdout"
        },
        {
          "output_type": "stream",
          "text": [
            "/usr/local/lib/python3.6/dist-packages/ipykernel_launcher.py:3: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
            "  This is separate from the ipykernel package so we can avoid doing imports until\n"
          ],
          "name": "stderr"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "oNiREwDe8Z1Q",
        "colab_type": "text"
      },
      "source": [
        "# 定义模型"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "8Bw0racHaM68",
        "colab_type": "text"
      },
      "source": [
        "## aquash 函数"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "V5MJV1tsSGbm",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "epsilon = 0.00000001\n",
        "def squash(x):\n",
        "    # not concern batch_size, maybe rewrite\n",
        "    s_squared_norm = torch.sum(x*x,1,keepdim=True) + epsilon\n",
        "    scale = torch.sqrt(s_squared_norm)/(1. + s_squared_norm)\n",
        "    # out = (batch_size,1,10)*(batch_size,16,10) = (batch_size,16,10)\n",
        "    out = scale * x\n",
        "    return out"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5nO6TgtFah8B",
        "colab_type": "text"
      },
      "source": [
        "## Capsule"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "lrC0IM4tahUA",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "class Capsule(nn.Module):\n",
        "\n",
        "    def __init__(self, in_units,in_channels, num_capsule, dim_capsule, routings=3, **kwargs):\n",
        "        super(Capsule, self).__init__(**kwargs)\n",
        "        self.in_units = in_units\n",
        "        self.in_channels = in_channels\n",
        "        self.num_capsule = num_capsule\n",
        "        self.dim_capsule = dim_capsule\n",
        "        self.routings = routings\n",
        "        # (in_units,10,128,16)\n",
        "        self.W = nn.Parameter((torch.randn(self.in_units,self.num_capsule,self.in_channels, self.dim_capsule)))\n",
        "\n",
        "    def forward(self, u_vecs):\n",
        "        u_vecs = u_vecs.permute(0,2,1)\n",
        "        u_vecs = u_vecs.unsqueeze(2)\n",
        "        u_vecs = u_vecs.unsqueeze(2)\n",
        "\n",
        "        \n",
        "        # (batch_size,in_units,1,1,in_channels)*(in_units,10,in_channels,16) = (batch_size,in_units,10,1,16)\n",
        "        u_hat_vecs = torch.matmul(u_vecs,self.W)\n",
        "        # (batch_size,in_units,10,16)\n",
        "        u_hat_vecs = u_hat_vecs.permute(0,1,2,4,3).squeeze(4)\n",
        "        \n",
        "        # (batch_size,10,in_units,16)\n",
        "        u_hat_vecs2 = u_hat_vecs.permute(0,2,1,3)\n",
        "    \n",
        "        # (batch_size,10,1,in_units)\n",
        "        b = torch.zeros(u_hat_vecs.size(0),self.num_capsule,1,self.in_units,device=DEVICE)\n",
        "        for i in range(self.routings):\n",
        "            # (batch_size,10,1,in_units)\n",
        "            c = F.softmax(b,-1)\n",
        "            # s = (batch_size,10,1,in_units)*(batch_size,10,in_units,16) = (batch_size,10,1,16)\n",
        "            s = torch.matmul(c,u_hat_vecs2)\n",
        "            # (batch_size,16,10)\n",
        "            s = s.permute(0,3,1,2).squeeze(3)\n",
        "            # (batch_size,16,10)\n",
        "            v = squash(s)\n",
        "            # here\n",
        "            # (batch_size,10,16,1)\n",
        "            v = v.permute(0,2,1).unsqueeze(3)\n",
        "            # (batch_size,10,in_units,16)*(batch_size,10,16,1) = (batch_size,10,in_units,1)\n",
        "            sim = torch.matmul(u_hat_vecs2,v)\n",
        "            # (batch_size,10,1,in_units)\n",
        "            sim = sim.permute(0,1,3,2)\n",
        "            b = b+sim\n",
        "        # (batch_size,16,10)\n",
        "        return v.permute(0,2,1,3).squeeze(3)"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "UFWNqnKYapTc",
        "colab_type": "text"
      },
      "source": [
        "## Dcoder（重构模型）"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "NSUYV_HtapwQ",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "# is_overlap = False\n",
        "class Decoder(nn.Module):\n",
        "    def __init__(self):\n",
        "        super(Decoder, self).__init__()\n",
        "        \n",
        "        self.reconstraction_layers = nn.Sequential(\n",
        "            nn.Linear(16 * 10, 512),\n",
        "            nn.ReLU(inplace=True),\n",
        "            nn.Linear(512, 1024),\n",
        "            nn.ReLU(inplace=True),\n",
        "            nn.Linear(1024, 784),\n",
        "            nn.Sigmoid()\n",
        "        )\n",
        "        \n",
        "    def forward(self, x,label):\n",
        "        # x (batch_size,10,16)\n",
        "        # mask (batch_size,10,1)\n",
        "        mask = label.unsqueeze(2)\n",
        "\n",
        "        inp = mask*x\n",
        "        inp = inp.view(x.size(0), -1)\n",
        "        reconstructions = self.reconstraction_layers(inp)\n",
        "        reconstructions = reconstructions.view(-1, 1, 28, 28)\n",
        "        \n",
        "        return reconstructions    "
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "89fn6iiga02f",
        "colab_type": "text"
      },
      "source": [
        "# CapsuleNet"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "tzVC4dbZ8Z1R",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "\n",
        "class CapsuleNet(nn.Module):\n",
        "    def __init__(self):\n",
        "        super().__init__()\n",
        "        # 1,28x28\n",
        "        self.conv1=nn.Conv2d(1,256,9) \n",
        "        self.conv2=nn.Conv2d(256,32*8,9,2) \n",
        "        self.capsule = Capsule(1152,8, 10, 16)\n",
        "        self.Decoder = Decoder()\n",
        "   \n",
        "    def forward(self,x,label):\n",
        "        batch_size = x.size(0)\n",
        "        # Conv1\n",
        "        out = self.conv1(x)\n",
        "        out = F.relu(out)\n",
        "        # PrimaryCaps\n",
        "        out = self.conv2(out)\n",
        "        # (batch_size, 32*8, 6, 6)\n",
        "        out = F.relu(out)\n",
        "        out = out.view(batch_size,8,-1)\n",
        "        out = squash(out)\n",
        "        # (batch_size,8,1152)\n",
        "        out = out.view(out.size(0),out.size(1),-1)\n",
        "        # Capsule\n",
        "        # (batch_size,16,10)\n",
        "        out = self.capsule(out)\n",
        "        # (batch_size,10,16)\n",
        "        out = out.permute(0,2,1)\n",
        "        # Decoder\n",
        "        decoder = self.Decoder(out,label)\n",
        "\n",
        "        \n",
        "\n",
        "\n",
        "        return out,decoder"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "56-Et_H_YM17",
        "colab_type": "text"
      },
      "source": [
        "# train\n",
        "已有训练好的参数，可跳过，直接执行后面的重构"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "WU6uRB2f9pFE",
        "colab_type": "code",
        "outputId": "5684f7a2-cdf9-4afd-849d-23ed988e74ba",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 290
        }
      },
      "source": [
        "import os\n",
        "model = CapsuleNet()\n",
        "if os.path.exists('params.pkl'):\n",
        "    model.load_state_dict(torch.load('params.pkl'))\n",
        "    \n",
        "model.to(DEVICE)"
      ],
      "execution_count": 12,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "CapsuleNet(\n",
              "  (conv1): Conv2d(1, 256, kernel_size=(9, 9), stride=(1, 1))\n",
              "  (conv2): Conv2d(256, 256, kernel_size=(9, 9), stride=(2, 2))\n",
              "  (capsule): Capsule()\n",
              "  (Decoder): Decoder(\n",
              "    (reconstraction_layers): Sequential(\n",
              "      (0): Linear(in_features=160, out_features=512, bias=True)\n",
              "      (1): ReLU(inplace)\n",
              "      (2): Linear(in_features=512, out_features=1024, bias=True)\n",
              "      (3): ReLU(inplace)\n",
              "      (4): Linear(in_features=1024, out_features=784, bias=True)\n",
              "      (5): Sigmoid()\n",
              "    )\n",
              "  )\n",
              ")"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 12
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "e7yBV9lN6I6C",
        "colab_type": "code",
        "outputId": "7ba3897e-d2a6-41cd-8358-9ee0b53cf4d4",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 2783
        }
      },
      "source": [
        "optimizer = optim.Adam(model.parameters())#, lr=0.000001)\n",
        "\n",
        "n_epoch = 30\n",
        "pre_acc = 0\n",
        "for epoch in range(n_epoch):\n",
        "    for batch_idx, (data, target) in enumerate(train_loader):\n",
        "        target = torch.sparse.torch.eye(10).index_select(dim=0, index=target)\n",
        "        data, target = data.to(DEVICE), target.to(DEVICE)\n",
        "        \n",
        "        optimizer.zero_grad()\n",
        "        \n",
        "        output,pred_img = model(data,target)\n",
        "        \n",
        "        # (batch_size,10)\n",
        "        output = torch.sqrt(torch.sum(output*output, 2))\n",
        "        loss1 = target*F.relu(0.9-output)**2 + 0.25*(1-target)*F.relu(output-0.1)**2\n",
        "        loss1 = loss1.sum(dim=1).mean()\n",
        "        loss2 = ((data-pred_img)**2).mean()\n",
        "        loss = loss1+0.0005*loss2\n",
        "        loss.backward()\n",
        "\n",
        "        optimizer.step()\n",
        "\n",
        "        if(batch_idx)%100 == 0:\n",
        "            y_true = torch.sparse.torch.eye(10).index_select(dim=0, index=y_test.cpu().data)\n",
        "            y_true = y_true.to(DEVICE)\n",
        "            predict,_ = model(x_test,y_true)\n",
        "            predict = torch.sqrt(torch.sum(predict*predict, 2))\n",
        "            _,pred = torch.max(predict, 1)\n",
        "\n",
        "            acc = torch.mean((torch.tensor(pred == y_test,dtype=torch.float)))\n",
        "\n",
        "            print('Train Epoch: {} [{}/{} ({:.0f}%)]\\tLoss1: {:.6f} \\tLoss2: {:.6f} \\t test accuracy:{:.4f}'.format(\n",
        "                epoch, batch_idx * len(data), len(train_loader.dataset),\n",
        "                100. * batch_idx / len(train_loader),loss1.item(),loss2.item(),acc))\n",
        "\n"
      ],
      "execution_count": 13,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "/usr/local/lib/python3.6/dist-packages/ipykernel_launcher.py:31: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n"
          ],
          "name": "stderr"
        },
        {
          "output_type": "stream",
          "text": [
            "Train Epoch: 0 [0/60000 (0%)]\tLoss1: 0.807812 \tLoss2: 1.229976 \t test accuracy:0.2734\n",
            "Train Epoch: 0 [12800/60000 (21%)]\tLoss1: 0.050368 \tLoss2: 0.798234 \t test accuracy:0.9531\n",
            "Train Epoch: 0 [25600/60000 (43%)]\tLoss1: 0.027365 \tLoss2: 0.794739 \t test accuracy:0.9844\n",
            "Train Epoch: 0 [38400/60000 (64%)]\tLoss1: 0.028302 \tLoss2: 0.755641 \t test accuracy:0.9922\n",
            "Train Epoch: 0 [51200/60000 (85%)]\tLoss1: 0.015003 \tLoss2: 0.716659 \t test accuracy:0.9922\n",
            "Train Epoch: 1 [0/60000 (0%)]\tLoss1: 0.010347 \tLoss2: 0.735337 \t test accuracy:0.9922\n",
            "Train Epoch: 1 [12800/60000 (21%)]\tLoss1: 0.007604 \tLoss2: 0.706147 \t test accuracy:0.9922\n",
            "Train Epoch: 1 [25600/60000 (43%)]\tLoss1: 0.023455 \tLoss2: 0.730091 \t test accuracy:0.9922\n",
            "Train Epoch: 1 [38400/60000 (64%)]\tLoss1: 0.009346 \tLoss2: 0.727677 \t test accuracy:0.9922\n",
            "Train Epoch: 1 [51200/60000 (85%)]\tLoss1: 0.005641 \tLoss2: 0.709103 \t test accuracy:1.0000\n",
            "Train Epoch: 2 [0/60000 (0%)]\tLoss1: 0.007930 \tLoss2: 0.733506 \t test accuracy:1.0000\n",
            "Train Epoch: 2 [12800/60000 (21%)]\tLoss1: 0.008013 \tLoss2: 0.708626 \t test accuracy:1.0000\n",
            "Train Epoch: 2 [25600/60000 (43%)]\tLoss1: 0.016147 \tLoss2: 0.733692 \t test accuracy:1.0000\n",
            "Train Epoch: 2 [38400/60000 (64%)]\tLoss1: 0.001661 \tLoss2: 0.722146 \t test accuracy:0.9922\n",
            "Train Epoch: 2 [51200/60000 (85%)]\tLoss1: 0.011580 \tLoss2: 0.692510 \t test accuracy:0.9922\n",
            "Train Epoch: 3 [0/60000 (0%)]\tLoss1: 0.009999 \tLoss2: 0.709514 \t test accuracy:1.0000\n",
            "Train Epoch: 3 [12800/60000 (21%)]\tLoss1: 0.002652 \tLoss2: 0.702732 \t test accuracy:1.0000\n",
            "Train Epoch: 3 [25600/60000 (43%)]\tLoss1: 0.001338 \tLoss2: 0.706783 \t test accuracy:0.9922\n",
            "Train Epoch: 3 [38400/60000 (64%)]\tLoss1: 0.004722 \tLoss2: 0.728133 \t test accuracy:1.0000\n",
            "Train Epoch: 3 [51200/60000 (85%)]\tLoss1: 0.003314 \tLoss2: 0.745347 \t test accuracy:1.0000\n",
            "Train Epoch: 4 [0/60000 (0%)]\tLoss1: 0.001176 \tLoss2: 0.725600 \t test accuracy:1.0000\n",
            "Train Epoch: 4 [12800/60000 (21%)]\tLoss1: 0.001371 \tLoss2: 0.724005 \t test accuracy:1.0000\n",
            "Train Epoch: 4 [25600/60000 (43%)]\tLoss1: 0.009000 \tLoss2: 0.729991 \t test accuracy:1.0000\n",
            "Train Epoch: 4 [38400/60000 (64%)]\tLoss1: 0.002560 \tLoss2: 0.740359 \t test accuracy:1.0000\n",
            "Train Epoch: 4 [51200/60000 (85%)]\tLoss1: 0.005670 \tLoss2: 0.717222 \t test accuracy:1.0000\n",
            "Train Epoch: 5 [0/60000 (0%)]\tLoss1: 0.002915 \tLoss2: 0.708905 \t test accuracy:1.0000\n",
            "Train Epoch: 5 [12800/60000 (21%)]\tLoss1: 0.004790 \tLoss2: 0.715563 \t test accuracy:1.0000\n",
            "Train Epoch: 5 [25600/60000 (43%)]\tLoss1: 0.000818 \tLoss2: 0.681266 \t test accuracy:1.0000\n",
            "Train Epoch: 5 [38400/60000 (64%)]\tLoss1: 0.003985 \tLoss2: 0.702296 \t test accuracy:1.0000\n",
            "Train Epoch: 5 [51200/60000 (85%)]\tLoss1: 0.001069 \tLoss2: 0.696817 \t test accuracy:1.0000\n",
            "Train Epoch: 6 [0/60000 (0%)]\tLoss1: 0.000289 \tLoss2: 0.713961 \t test accuracy:1.0000\n",
            "Train Epoch: 6 [12800/60000 (21%)]\tLoss1: 0.000468 \tLoss2: 0.715397 \t test accuracy:1.0000\n",
            "Train Epoch: 6 [25600/60000 (43%)]\tLoss1: 0.000888 \tLoss2: 0.714441 \t test accuracy:1.0000\n",
            "Train Epoch: 6 [38400/60000 (64%)]\tLoss1: 0.003041 \tLoss2: 0.719686 \t test accuracy:1.0000\n",
            "Train Epoch: 6 [51200/60000 (85%)]\tLoss1: 0.001312 \tLoss2: 0.723361 \t test accuracy:1.0000\n",
            "Train Epoch: 7 [0/60000 (0%)]\tLoss1: 0.002312 \tLoss2: 0.716397 \t test accuracy:1.0000\n",
            "Train Epoch: 7 [12800/60000 (21%)]\tLoss1: 0.002540 \tLoss2: 0.710901 \t test accuracy:1.0000\n",
            "Train Epoch: 7 [25600/60000 (43%)]\tLoss1: 0.000487 \tLoss2: 0.712496 \t test accuracy:1.0000\n",
            "Train Epoch: 7 [38400/60000 (64%)]\tLoss1: 0.000840 \tLoss2: 0.690105 \t test accuracy:0.9922\n",
            "Train Epoch: 7 [51200/60000 (85%)]\tLoss1: 0.000227 \tLoss2: 0.681407 \t test accuracy:1.0000\n",
            "Train Epoch: 8 [0/60000 (0%)]\tLoss1: 0.000197 \tLoss2: 0.714004 \t test accuracy:1.0000\n",
            "Train Epoch: 8 [12800/60000 (21%)]\tLoss1: 0.001634 \tLoss2: 0.691610 \t test accuracy:1.0000\n",
            "Train Epoch: 8 [25600/60000 (43%)]\tLoss1: 0.000360 \tLoss2: 0.683666 \t test accuracy:1.0000\n",
            "Train Epoch: 8 [38400/60000 (64%)]\tLoss1: 0.001890 \tLoss2: 0.696268 \t test accuracy:1.0000\n",
            "Train Epoch: 8 [51200/60000 (85%)]\tLoss1: 0.000528 \tLoss2: 0.678428 \t test accuracy:1.0000\n",
            "Train Epoch: 9 [0/60000 (0%)]\tLoss1: 0.001936 \tLoss2: 0.693992 \t test accuracy:1.0000\n",
            "Train Epoch: 9 [12800/60000 (21%)]\tLoss1: 0.000533 \tLoss2: 0.705128 \t test accuracy:1.0000\n",
            "Train Epoch: 9 [25600/60000 (43%)]\tLoss1: 0.000183 \tLoss2: 0.665193 \t test accuracy:0.9922\n",
            "Train Epoch: 9 [38400/60000 (64%)]\tLoss1: 0.001472 \tLoss2: 0.696482 \t test accuracy:1.0000\n",
            "Train Epoch: 9 [51200/60000 (85%)]\tLoss1: 0.000915 \tLoss2: 0.693081 \t test accuracy:1.0000\n",
            "Train Epoch: 10 [0/60000 (0%)]\tLoss1: 0.000523 \tLoss2: 0.694607 \t test accuracy:1.0000\n",
            "Train Epoch: 10 [12800/60000 (21%)]\tLoss1: 0.000455 \tLoss2: 0.687156 \t test accuracy:0.9922\n",
            "Train Epoch: 10 [25600/60000 (43%)]\tLoss1: 0.000080 \tLoss2: 0.719835 \t test accuracy:1.0000\n",
            "Train Epoch: 10 [38400/60000 (64%)]\tLoss1: 0.002280 \tLoss2: 0.669596 \t test accuracy:1.0000\n",
            "Train Epoch: 10 [51200/60000 (85%)]\tLoss1: 0.000220 \tLoss2: 0.701907 \t test accuracy:1.0000\n",
            "Train Epoch: 11 [0/60000 (0%)]\tLoss1: 0.000707 \tLoss2: 0.696027 \t test accuracy:1.0000\n",
            "Train Epoch: 11 [12800/60000 (21%)]\tLoss1: 0.000187 \tLoss2: 0.694867 \t test accuracy:1.0000\n",
            "Train Epoch: 11 [25600/60000 (43%)]\tLoss1: 0.001261 \tLoss2: 0.714062 \t test accuracy:1.0000\n",
            "Train Epoch: 11 [38400/60000 (64%)]\tLoss1: 0.001479 \tLoss2: 0.690665 \t test accuracy:0.9922\n",
            "Train Epoch: 11 [51200/60000 (85%)]\tLoss1: 0.000257 \tLoss2: 0.705418 \t test accuracy:1.0000\n",
            "Train Epoch: 12 [0/60000 (0%)]\tLoss1: 0.004758 \tLoss2: 0.711279 \t test accuracy:1.0000\n",
            "Train Epoch: 12 [12800/60000 (21%)]\tLoss1: 0.001642 \tLoss2: 0.702149 \t test accuracy:1.0000\n",
            "Train Epoch: 12 [25600/60000 (43%)]\tLoss1: 0.000100 \tLoss2: 0.684998 \t test accuracy:1.0000\n",
            "Train Epoch: 12 [38400/60000 (64%)]\tLoss1: 0.000565 \tLoss2: 0.706967 \t test accuracy:1.0000\n",
            "Train Epoch: 12 [51200/60000 (85%)]\tLoss1: 0.000253 \tLoss2: 0.707027 \t test accuracy:1.0000\n",
            "Train Epoch: 13 [0/60000 (0%)]\tLoss1: 0.000421 \tLoss2: 0.700813 \t test accuracy:1.0000\n",
            "Train Epoch: 13 [12800/60000 (21%)]\tLoss1: 0.000120 \tLoss2: 0.685743 \t test accuracy:1.0000\n",
            "Train Epoch: 13 [25600/60000 (43%)]\tLoss1: 0.000218 \tLoss2: 0.703059 \t test accuracy:1.0000\n",
            "Train Epoch: 13 [38400/60000 (64%)]\tLoss1: 0.001207 \tLoss2: 0.717006 \t test accuracy:1.0000\n",
            "Train Epoch: 13 [51200/60000 (85%)]\tLoss1: 0.000354 \tLoss2: 0.695582 \t test accuracy:1.0000\n",
            "Train Epoch: 14 [0/60000 (0%)]\tLoss1: 0.000279 \tLoss2: 0.694241 \t test accuracy:1.0000\n",
            "Train Epoch: 14 [12800/60000 (21%)]\tLoss1: 0.000048 \tLoss2: 0.713103 \t test accuracy:1.0000\n",
            "Train Epoch: 14 [25600/60000 (43%)]\tLoss1: 0.000153 \tLoss2: 0.696319 \t test accuracy:1.0000\n",
            "Train Epoch: 14 [38400/60000 (64%)]\tLoss1: 0.000106 \tLoss2: 0.685855 \t test accuracy:1.0000\n",
            "Train Epoch: 14 [51200/60000 (85%)]\tLoss1: 0.001474 \tLoss2: 0.676437 \t test accuracy:1.0000\n",
            "Train Epoch: 15 [0/60000 (0%)]\tLoss1: 0.002064 \tLoss2: 0.695848 \t test accuracy:1.0000\n",
            "Train Epoch: 15 [12800/60000 (21%)]\tLoss1: 0.000612 \tLoss2: 0.684360 \t test accuracy:1.0000\n",
            "Train Epoch: 15 [25600/60000 (43%)]\tLoss1: 0.001461 \tLoss2: 0.681717 \t test accuracy:1.0000\n",
            "Train Epoch: 15 [38400/60000 (64%)]\tLoss1: 0.001145 \tLoss2: 0.682072 \t test accuracy:1.0000\n",
            "Train Epoch: 15 [51200/60000 (85%)]\tLoss1: 0.001662 \tLoss2: 0.664362 \t test accuracy:0.9922\n",
            "Train Epoch: 16 [0/60000 (0%)]\tLoss1: 0.001335 \tLoss2: 0.722690 \t test accuracy:1.0000\n",
            "Train Epoch: 16 [12800/60000 (21%)]\tLoss1: 0.000251 \tLoss2: 0.695980 \t test accuracy:1.0000\n",
            "Train Epoch: 16 [25600/60000 (43%)]\tLoss1: 0.000244 \tLoss2: 0.682645 \t test accuracy:1.0000\n",
            "Train Epoch: 16 [38400/60000 (64%)]\tLoss1: 0.000836 \tLoss2: 0.668376 \t test accuracy:1.0000\n",
            "Train Epoch: 16 [51200/60000 (85%)]\tLoss1: 0.000086 \tLoss2: 0.662689 \t test accuracy:1.0000\n",
            "Train Epoch: 17 [0/60000 (0%)]\tLoss1: 0.000135 \tLoss2: 0.686563 \t test accuracy:1.0000\n",
            "Train Epoch: 17 [12800/60000 (21%)]\tLoss1: 0.000107 \tLoss2: 0.681677 \t test accuracy:1.0000\n",
            "Train Epoch: 17 [25600/60000 (43%)]\tLoss1: 0.000527 \tLoss2: 0.678590 \t test accuracy:1.0000\n",
            "Train Epoch: 17 [38400/60000 (64%)]\tLoss1: 0.000079 \tLoss2: 0.705296 \t test accuracy:1.0000\n",
            "Train Epoch: 17 [51200/60000 (85%)]\tLoss1: 0.000214 \tLoss2: 0.705435 \t test accuracy:1.0000\n",
            "Train Epoch: 18 [0/60000 (0%)]\tLoss1: 0.000045 \tLoss2: 0.656969 \t test accuracy:1.0000\n",
            "Train Epoch: 18 [12800/60000 (21%)]\tLoss1: 0.000022 \tLoss2: 0.661251 \t test accuracy:1.0000\n",
            "Train Epoch: 18 [25600/60000 (43%)]\tLoss1: 0.000063 \tLoss2: 0.696687 \t test accuracy:1.0000\n",
            "Train Epoch: 18 [38400/60000 (64%)]\tLoss1: 0.000259 \tLoss2: 0.684626 \t test accuracy:1.0000\n",
            "Train Epoch: 18 [51200/60000 (85%)]\tLoss1: 0.000594 \tLoss2: 0.678900 \t test accuracy:1.0000\n",
            "Train Epoch: 19 [0/60000 (0%)]\tLoss1: 0.000025 \tLoss2: 0.675741 \t test accuracy:1.0000\n",
            "Train Epoch: 19 [12800/60000 (21%)]\tLoss1: 0.000412 \tLoss2: 0.718886 \t test accuracy:0.9922\n",
            "Train Epoch: 19 [25600/60000 (43%)]\tLoss1: 0.000054 \tLoss2: 0.701096 \t test accuracy:1.0000\n",
            "Train Epoch: 19 [38400/60000 (64%)]\tLoss1: 0.000191 \tLoss2: 0.646564 \t test accuracy:1.0000\n",
            "Train Epoch: 19 [51200/60000 (85%)]\tLoss1: 0.000013 \tLoss2: 0.662314 \t test accuracy:1.0000\n",
            "Train Epoch: 20 [0/60000 (0%)]\tLoss1: 0.000576 \tLoss2: 0.693048 \t test accuracy:1.0000\n",
            "Train Epoch: 20 [12800/60000 (21%)]\tLoss1: 0.000324 \tLoss2: 0.668652 \t test accuracy:1.0000\n",
            "Train Epoch: 20 [25600/60000 (43%)]\tLoss1: 0.000228 \tLoss2: 0.657819 \t test accuracy:1.0000\n",
            "Train Epoch: 20 [38400/60000 (64%)]\tLoss1: 0.000184 \tLoss2: 0.685319 \t test accuracy:0.9922\n",
            "Train Epoch: 20 [51200/60000 (85%)]\tLoss1: 0.000245 \tLoss2: 0.705876 \t test accuracy:1.0000\n",
            "Train Epoch: 21 [0/60000 (0%)]\tLoss1: 0.001037 \tLoss2: 0.675474 \t test accuracy:1.0000\n",
            "Train Epoch: 21 [12800/60000 (21%)]\tLoss1: 0.000933 \tLoss2: 0.690558 \t test accuracy:1.0000\n",
            "Train Epoch: 21 [25600/60000 (43%)]\tLoss1: 0.000064 \tLoss2: 0.661630 \t test accuracy:1.0000\n",
            "Train Epoch: 21 [38400/60000 (64%)]\tLoss1: 0.000013 \tLoss2: 0.705780 \t test accuracy:1.0000\n",
            "Train Epoch: 21 [51200/60000 (85%)]\tLoss1: 0.000266 \tLoss2: 0.673887 \t test accuracy:0.9922\n",
            "Train Epoch: 22 [0/60000 (0%)]\tLoss1: 0.000143 \tLoss2: 0.677331 \t test accuracy:1.0000\n",
            "Train Epoch: 22 [12800/60000 (21%)]\tLoss1: 0.000015 \tLoss2: 0.676061 \t test accuracy:1.0000\n",
            "Train Epoch: 22 [25600/60000 (43%)]\tLoss1: 0.000042 \tLoss2: 0.678370 \t test accuracy:1.0000\n",
            "Train Epoch: 22 [38400/60000 (64%)]\tLoss1: 0.000035 \tLoss2: 0.672825 \t test accuracy:1.0000\n",
            "Train Epoch: 22 [51200/60000 (85%)]\tLoss1: 0.000037 \tLoss2: 0.694205 \t test accuracy:1.0000\n",
            "Train Epoch: 23 [0/60000 (0%)]\tLoss1: 0.001625 \tLoss2: 0.668901 \t test accuracy:1.0000\n",
            "Train Epoch: 23 [12800/60000 (21%)]\tLoss1: 0.000120 \tLoss2: 0.655298 \t test accuracy:1.0000\n",
            "Train Epoch: 23 [25600/60000 (43%)]\tLoss1: 0.000006 \tLoss2: 0.676081 \t test accuracy:1.0000\n",
            "Train Epoch: 23 [38400/60000 (64%)]\tLoss1: 0.000311 \tLoss2: 0.653579 \t test accuracy:1.0000\n",
            "Train Epoch: 23 [51200/60000 (85%)]\tLoss1: 0.000337 \tLoss2: 0.656271 \t test accuracy:0.9922\n",
            "Train Epoch: 24 [0/60000 (0%)]\tLoss1: 0.000051 \tLoss2: 0.690603 \t test accuracy:0.9922\n",
            "Train Epoch: 24 [12800/60000 (21%)]\tLoss1: 0.000094 \tLoss2: 0.655645 \t test accuracy:0.9922\n",
            "Train Epoch: 24 [25600/60000 (43%)]\tLoss1: 0.000382 \tLoss2: 0.686946 \t test accuracy:1.0000\n",
            "Train Epoch: 24 [38400/60000 (64%)]\tLoss1: 0.000312 \tLoss2: 0.661412 \t test accuracy:1.0000\n",
            "Train Epoch: 24 [51200/60000 (85%)]\tLoss1: 0.000015 \tLoss2: 0.680440 \t test accuracy:0.9922\n",
            "Train Epoch: 25 [0/60000 (0%)]\tLoss1: 0.000525 \tLoss2: 0.690955 \t test accuracy:1.0000\n",
            "Train Epoch: 25 [12800/60000 (21%)]\tLoss1: 0.000522 \tLoss2: 0.654996 \t test accuracy:1.0000\n",
            "Train Epoch: 25 [25600/60000 (43%)]\tLoss1: 0.000331 \tLoss2: 0.668230 \t test accuracy:1.0000\n",
            "Train Epoch: 25 [38400/60000 (64%)]\tLoss1: 0.000066 \tLoss2: 0.697684 \t test accuracy:0.9922\n",
            "Train Epoch: 25 [51200/60000 (85%)]\tLoss1: 0.000492 \tLoss2: 0.671946 \t test accuracy:1.0000\n",
            "Train Epoch: 26 [0/60000 (0%)]\tLoss1: 0.000340 \tLoss2: 0.693028 \t test accuracy:1.0000\n",
            "Train Epoch: 26 [12800/60000 (21%)]\tLoss1: 0.000019 \tLoss2: 0.671784 \t test accuracy:1.0000\n",
            "Train Epoch: 26 [25600/60000 (43%)]\tLoss1: 0.000092 \tLoss2: 0.684917 \t test accuracy:1.0000\n",
            "Train Epoch: 26 [38400/60000 (64%)]\tLoss1: 0.000027 \tLoss2: 0.658519 \t test accuracy:1.0000\n",
            "Train Epoch: 26 [51200/60000 (85%)]\tLoss1: 0.000051 \tLoss2: 0.662214 \t test accuracy:1.0000\n",
            "Train Epoch: 27 [0/60000 (0%)]\tLoss1: 0.000060 \tLoss2: 0.680690 \t test accuracy:1.0000\n",
            "Train Epoch: 27 [12800/60000 (21%)]\tLoss1: 0.000202 \tLoss2: 0.662924 \t test accuracy:1.0000\n",
            "Train Epoch: 27 [25600/60000 (43%)]\tLoss1: 0.000091 \tLoss2: 0.688571 \t test accuracy:1.0000\n",
            "Train Epoch: 27 [38400/60000 (64%)]\tLoss1: 0.000029 \tLoss2: 0.673507 \t test accuracy:1.0000\n",
            "Train Epoch: 27 [51200/60000 (85%)]\tLoss1: 0.000012 \tLoss2: 0.662414 \t test accuracy:1.0000\n",
            "Train Epoch: 28 [0/60000 (0%)]\tLoss1: 0.000068 \tLoss2: 0.661478 \t test accuracy:1.0000\n",
            "Train Epoch: 28 [12800/60000 (21%)]\tLoss1: 0.000069 \tLoss2: 0.676011 \t test accuracy:0.9922\n",
            "Train Epoch: 28 [25600/60000 (43%)]\tLoss1: 0.000085 \tLoss2: 0.655647 \t test accuracy:1.0000\n",
            "Train Epoch: 28 [38400/60000 (64%)]\tLoss1: 0.000129 \tLoss2: 0.645794 \t test accuracy:1.0000\n",
            "Train Epoch: 28 [51200/60000 (85%)]\tLoss1: 0.000035 \tLoss2: 0.680968 \t test accuracy:1.0000\n",
            "Train Epoch: 29 [0/60000 (0%)]\tLoss1: 0.000043 \tLoss2: 0.679005 \t test accuracy:1.0000\n",
            "Train Epoch: 29 [12800/60000 (21%)]\tLoss1: 0.000118 \tLoss2: 0.673500 \t test accuracy:1.0000\n",
            "Train Epoch: 29 [25600/60000 (43%)]\tLoss1: 0.000033 \tLoss2: 0.662643 \t test accuracy:0.9922\n",
            "Train Epoch: 29 [38400/60000 (64%)]\tLoss1: 0.000569 \tLoss2: 0.650962 \t test accuracy:1.0000\n",
            "Train Epoch: 29 [51200/60000 (85%)]\tLoss1: 0.000028 \tLoss2: 0.689124 \t test accuracy:1.0000\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "RWLP4KQgeRTj",
        "colab_type": "text"
      },
      "source": [
        "# 计算测试集的准确度\n",
        "之前训练时只用了一个batch_size，现在计算所有"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "zmyOA3iQdZlP",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 92
        },
        "outputId": "5d7ab670-6f9b-4352-ec61-f852c0bd861e"
      },
      "source": [
        "test_acc = []\n",
        "for batch_idx, (data, target) in enumerate(test_loader):\n",
        "    data, target = data.to(DEVICE), target.to(DEVICE)\n",
        "    y_true = torch.sparse.torch.eye(10).index_select(dim=0, index=target.cpu().data)\n",
        "    y_true = y_true.to(DEVICE)\n",
        "    predict,_ = model(data,y_true)\n",
        "    predict = torch.sqrt(torch.sum(predict*predict, 2))\n",
        "    _,pred = torch.max(predict, 1)\n",
        "\n",
        "    acc = torch.mean((torch.tensor(pred == target,dtype=torch.float)))\n",
        "    test_acc.append(acc)\n",
        "print('total test accuracy : %.4f'%(np.array(test_acc).mean()))"
      ],
      "execution_count": 81,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "/usr/local/lib/python3.6/dist-packages/ipykernel_launcher.py:10: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).\n",
            "  # Remove the CWD from sys.path while we load stuff.\n"
          ],
          "name": "stderr"
        },
        {
          "output_type": "stream",
          "text": [
            "total test accuracy : 0.9924\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "G440jWVbOxNq",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        "torch.save(model.state_dict(), 'params.pkl')"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "2LUT6f4dYQaW",
        "colab_type": "text"
      },
      "source": [
        "# 加载训练好的模型"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "WQyWBkWoELzr",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 290
        },
        "outputId": "e66bf864-fd48-47f1-b16a-9d9acf20b315"
      },
      "source": [
        "import os\n",
        "model = CapsuleNet()\n",
        "model.load_state_dict(torch.load('params.pkl'))\n",
        "\n",
        "model.to(DEVICE)\n",
        "print(model)"
      ],
      "execution_count": 62,
      "outputs": [
        {
          "output_type": "stream",
          "text": [
            "CapsuleNet(\n",
            "  (conv1): Conv2d(1, 256, kernel_size=(9, 9), stride=(1, 1))\n",
            "  (conv2): Conv2d(256, 256, kernel_size=(9, 9), stride=(2, 2))\n",
            "  (capsule): Capsule()\n",
            "  (Decoder): Decoder(\n",
            "    (reconstraction_layers): Sequential(\n",
            "      (0): Linear(in_features=160, out_features=512, bias=True)\n",
            "      (1): ReLU(inplace)\n",
            "      (2): Linear(in_features=512, out_features=1024, bias=True)\n",
            "      (3): ReLU(inplace)\n",
            "      (4): Linear(in_features=1024, out_features=784, bias=True)\n",
            "      (5): Sigmoid()\n",
            "    )\n",
            "  )\n",
            ")\n"
          ],
          "name": "stdout"
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "IiEla9fw3U0n",
        "colab_type": "text"
      },
      "source": [
        "# 单数字重构"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "sPiqcyoh3EZv",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 405
        },
        "outputId": "8454ba70-6b1b-4a99-b6c3-936bc7a69fc2"
      },
      "source": [
        "import matplotlib.pyplot as plt\n",
        "import random\n",
        "import numpy as np\n",
        "import time\n",
        "\n",
        "img_num = 7\n",
        "k = random.randint(10,100)\n",
        "\n",
        "imgs = x_test[k:img_num+k].cpu().numpy()\n",
        "label = y_test[k:img_num+k].cpu().numpy()\n",
        "\n",
        "y_true = torch.sparse.torch.eye(10).index_select(dim=0, index=y_test[k:img_num+k].cpu().data)\n",
        "y_true = y_true.cuda()\n",
        "\n",
        "out,pred_imgs = model(x_test[k:img_num+k],y_true)\n",
        "pred_label = torch.sqrt(torch.sum(out*out, 2))\n",
        "\n",
        "pred_imgs = pred_imgs.cpu().detach().numpy()\n",
        "pred_label = pred_label.cpu().detach().numpy()\n",
        "pred_label = np.argmax(pred_label,1)\n",
        "\n",
        "plt.figure(\"test Image\",figsize=(img_num*3,3*2))\n",
        "for i in range(img_num):\n",
        "    index = i\n",
        "    img = imgs[index]\n",
        "    title = label[index]\n",
        "    img = img.reshape([28, 28])\n",
        "    \n",
        "    plt.subplot(2,img_num,i+1)\n",
        "    plt.imshow(img,cmap=plt.cm.gray)\n",
        "    plt.axis('off') \n",
        "    plt.title(title)\n",
        "    \n",
        "for i in range(img_num):\n",
        "    index = i\n",
        "    img = pred_imgs[index]\n",
        "    plt.subplot(2,img_num,i+img_num+1)\n",
        "    \n",
        "    title = pred_label[index]\n",
        "    img = img.reshape([28, 28])\n",
        "    plt.imshow(img,cmap=plt.cm.gray)\n",
        "    plt.axis('off') \n",
        "    plt.title(title)\n",
        "    \n",
        "filename = 'r1'+str(time.time())+'.jpg'\n",
        "plt.savefig(filename)\n",
        "plt.show()"
      ],
      "execution_count": 82,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAABLUAAAFwCAYAAABO2EmYAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
            "text/plain": [
              "<Figure size 1512x432 with 14 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "gQYN7bFIYau9",
        "colab_type": "text"
      },
      "source": [
        "# 2个数字重叠重构(目前还有问题)"
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "jexXL2AahsE3",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 670
        },
        "outputId": "10e64e60-8407-4aed-c77a-8307b176b627"
      },
      "source": [
        "import random\n",
        "k = random.randint(10,100)\n",
        "\n",
        "img_num = 6\n",
        "imgs = (torch.max(x_test[:img_num],x_test[k:img_num+k])).cpu().numpy()\n",
        "\n",
        "label1 = (y_test[:img_num]).cpu().numpy()\n",
        "label2 = (y_test[k:img_num+k]).cpu().numpy()\n",
        "\n",
        "y_true1 = torch.sparse.torch.eye(10).index_select(dim=0, index=y_test[:img_num].cpu().data)\n",
        "y_true1 = y_true1.to(DEVICE)\n",
        "\n",
        "y_true2 = torch.sparse.torch.eye(10).index_select(dim=0, index=y_test[k:img_num+k].cpu().data)\n",
        "y_true2 = y_true2.to(DEVICE)\n",
        "\n",
        "data = torch.tensor(imgs,dtype=torch.float,device=DEVICE)\n",
        "\n",
        "out1,pred_imgs1 = model(data,y_true1)\n",
        "out2,pred_imgs2 = model(data,y_true2)\n",
        "\n",
        "pred_label1 = torch.sqrt(torch.sum(out1*out1, 2))\n",
        "pred_label2 = pred_label1\n",
        "\n",
        "\n",
        "pred_imgs1 = pred_imgs1.cpu().detach().numpy()\n",
        "pred_imgs2 = pred_imgs2.cpu().detach().numpy()\n",
        "\n",
        "pred_label1 = pred_label1.cpu().detach().numpy()\n",
        "pred_label2 = pred_label2.cpu().detach().numpy()\n",
        "\n",
        "pred_label1 = np.argsort(pred_label1,1)[:,-1]\n",
        "pred_label2 =np.argsort(pred_label2,1)[:,-2]\n",
        "\n",
        "pred_imgs3 = np.zeros_like(pred_imgs1)\n",
        "\n",
        "pred_imgs = np.stack((pred_imgs1,pred_imgs2,pred_imgs3),1)\n",
        "pred_imgs = pred_imgs.transpose(0,3,4,1,2).squeeze(4)\n",
        "\n",
        "\n",
        "plt.figure(\"test Image\",figsize=(img_num*2,6*2))\n",
        "for i in range(img_num):\n",
        "    index = i\n",
        "    img = imgs[index]\n",
        "    title = str(label1[index])+\" \"+str(label2[index])\n",
        "    img = img.reshape([28, 28])\n",
        "    \n",
        "    plt.subplot(4,img_num,i+1)\n",
        "    plt.imshow(img,cmap=plt.cm.gray)\n",
        "    plt.axis('off') \n",
        "    plt.title(title)\n",
        "    \n",
        "for i in range(img_num):\n",
        "    index = i\n",
        "    \n",
        "    plt.subplot(4,img_num,i+img_num+1)\n",
        "    title = str(pred_label1[index])+\" \"+str(pred_label2[index])\n",
        "    \n",
        "    img = pred_imgs[index]\n",
        "    plt.imshow(img)\n",
        "    plt.axis('off') \n",
        "    plt.title(title)\n",
        "    \n",
        "    img1 = pred_imgs1[index]\n",
        "    img1 = img1.reshape([28, 28])    \n",
        "    plt.subplot(4,img_num,i+2*img_num+1)\n",
        "    plt.imshow(img1,cmap=plt.cm.gray)\n",
        "    plt.axis('off') \n",
        "    plt.title(title)\n",
        "    \n",
        "    img2 = pred_imgs2[index]\n",
        "    img2 = img2.reshape([28, 28])    \n",
        "    plt.subplot(4,img_num,i+3*img_num+1)\n",
        "    plt.imshow(img2,cmap=plt.cm.gray)\n",
        "    plt.axis('off') \n",
        "    plt.title(title)\n",
        "\n",
        "\n",
        "    \n",
        "# filename = 'r2'+str(time.time())+'.jpg'\n",
        "# plt.savefig(filename)    \n",
        "plt.show()"
      ],
      "execution_count": 83,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "image/png": "iVBORw0KGgoAAAANSUhEUgAAAr8AAAKNCAYAAAApqwNIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAABkg+QXAAAA2fj/NDskeV2XABgAAAAASUVORK5CYII=\n",
            "text/plain": [
              "<Figure size 864x864 with 24 Axes>"
            ]
          },
          "metadata": {
            "tags": []
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "j_XAjqROoB6j",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        ""
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "cell_type": "code",
      "metadata": {
        "id": "1kj2RqFx1nCM",
        "colab_type": "code",
        "colab": {}
      },
      "source": [
        ""
      ],
      "execution_count": 0,
      "outputs": []
    }
  ]
}