10_Introduction_to_Convolutional_Neural_Nets_workset.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "id": "VSlqSSC-1nAO"
   },
   "outputs": [],
   "source": [
    "import torch\n",
    "from torch import nn\n",
    "import torch.nn.functional as F\n",
    "import torchvision as tv\n",
    "\n",
    "from PIL import Image\n",
    "import requests\n",
    "from io import BytesIO\n",
    "\n",
    "import matplotlib.pyplot as plt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 417
    },
    "id": "wq_ch3ZZLYmH",
    "outputId": "53619a69-39fc-42f4-e7b9-9999544506f8"
   },
   "outputs": [],
   "source": [
    "url = 'https://media.istockphoto.com/vectors/chess-silhouettes-vector-id165635822?b=1&k=20&m=165635822&s=612x612&w=0&h=pmf6FVa--nzyWCKb0SyTkIi3xdaHaamJuaR-FIjw1iI='\n",
    "response = requests.get(url)\n",
    "img = Image.open(BytesIO(response.content)).convert('L')\n",
    "img"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "id": "KFlFdiI4S7J6"
   },
   "outputs": [],
   "source": [
    "to_tensor = tv.transforms.ToTensor()\n",
    "to_image = tv.transforms.ToPILImage()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "TjP50az6Rdnh",
    "outputId": "5125d276-ea5a-4a4c-a548-b2e21f9112e7"
   },
   "outputs": [],
   "source": [
    "img_tensor = (\n",
    "    to_tensor(img) # Convert to tensor\n",
    "    .unsqueeze(0) # Add a batch dimension\n",
    "    / 255. # Scale between 0 and 1.\n",
    ")\n",
    "img_tensor.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "id": "n7LA-TaGKhp2"
   },
   "outputs": [],
   "source": [
    "vertical_filter = torch.tensor([\n",
    "    [1., 0, -1],\n",
    "    [1., 0, -1],\n",
    "    [1., 0, -1],\n",
    "]).unsqueeze(0).unsqueeze(0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "id": "YI_ss_TqKHZg"
   },
   "outputs": [],
   "source": [
    "vert_output = F.conv2d(img_tensor, vertical_filter).squeeze()\n",
    "vert_image = to_image(vert_output)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 313
    },
    "id": "Xsyg5iTATCV4",
    "outputId": "4553b3a5-257e-48f6-8c09-1fd22719efdc"
   },
   "outputs": [],
   "source": [
    "fig, axes = plt.subplots(1, 2, figsize=(20, 10))\n",
    "axes[0].imshow(img, cmap='Greys_r')\n",
    "axes[0].set_title('Original greyscale image')\n",
    "axes[1].imshow(vert_image, cmap='Greys_r')\n",
    "axes[1].set_title('Feature map of vertical filter')\n",
    "for ax in axes:\n",
    "    ax.set_xticks([])\n",
    "    ax.set_yticks([])\n",
    "fig.tight_layout()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "anrR11WXu7_L"
   },
   "source": [
    "# Exercise 10.1\n",
    "\n",
    "Use what you learned in this lesson to create a horizontal feature map.\n",
    "The result of this exercise should be an edge detector that detects the top edges of the chess pieces and squares on the board.\n",
    "\n",
    "Then, use the `F.conv2d` to apply the filter.\n",
    "\n",
    "<!-- startquestion -->"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "id": "oC51Xvw8wWoD"
   },
   "outputs": [],
   "source": [
    "# Your code here\n",
    "horizontal_filter = ...\n",
    "\n",
    "horizontal_tensor = ...\n",
    "horizontal_image = ..."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "id": "2oWAWKCbwej1"
   },
   "outputs": [],
   "source": [
    "assert horizontal_filter.ndim == 4"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 313
    },
    "id": "nodTP1FNwA0F",
    "outputId": "b748eb9f-27f0-4791-db9c-af877edc10bb"
   },
   "outputs": [],
   "source": [
    "fig, axes = plt.subplots(1, 2, figsize=(20, 10))\n",
    "axes[0].imshow(img, cmap='Greys_r')\n",
    "axes[0].set_title('Original greyscale image')\n",
    "axes[1].imshow(horizontal_image, cmap='Greys_r')\n",
    "axes[1].set_title('Feature map of horizontal filter')\n",
    "for ax in axes:\n",
    "    ax.set_xticks([])\n",
    "    ax.set_yticks([])\n",
    "fig.tight_layout()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "-LSXV7pvxpiD",
    "outputId": "8bc541eb-f31e-4694-81cf-b4c6db6b5fbb"
   },
   "outputs": [],
   "source": [
    "two_filters = torch.cat([vertical_filter, horizontal_filter], dim=0)\n",
    "two_filters.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "jfTSAyM9zmEq",
    "outputId": "1d600559-22e8-4da1-b89e-f8eb73cd0712"
   },
   "outputs": [],
   "source": [
    "two_filter_output = F.conv2d(img_tensor, two_filters)\n",
    "two_filter_output.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "u-RdZJjrzz4V"
   },
   "outputs": [],
   "source": [
    "# unpack the filters into their own tensors\n",
    "vert_output, horizontal_output = two_filter_output.squeeze() \n",
    "# convert tensors to images\n",
    "vert_image = to_image(vert_output.squeeze())\n",
    "horizontal_image = to_image(horizontal_output.squeeze())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 507
    },
    "id": "GlQMBrYIz7KY",
    "outputId": "95c091e2-cd4a-465e-fe7f-df7beb601ebe"
   },
   "outputs": [],
   "source": [
    "fig, axes = plt.subplots(1, 2, figsize=(20, 10))\n",
    "axes[0].imshow(vert_image, cmap='Greys_r')\n",
    "axes[0].set_title('Feature map of vertical filter')\n",
    "axes[1].imshow(horizontal_image, cmap='Greys_r')\n",
    "axes[1].set_title('Feature map of horizontal filter')\n",
    "for ax in axes:\n",
    "    ax.set_xticks([])\n",
    "    ax.set_yticks([])\n",
    "fig.tight_layout()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "id": "SWacj1EV6MhK"
   },
   "outputs": [],
   "source": [
    "zero_pad = nn.ZeroPad2d(30)\n",
    "reflection_pad = nn.ReflectionPad2d(30)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 477
    },
    "id": "oXIHxf7eLmEZ",
    "outputId": "6455cc36-1f44-4481-8c9a-52175ec98031"
   },
   "outputs": [],
   "source": [
    "to_image(zero_pad(img_tensor).squeeze() * 255)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 477
    },
    "id": "XM-4XLtZ7WzT",
    "outputId": "d8f17b2f-be56-4dc1-8245-9b56b4fae764"
   },
   "outputs": [],
   "source": [
    "to_image(reflection_pad(img_tensor).squeeze() * 255)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "id": "k1rMd1aOO0bm"
   },
   "outputs": [],
   "source": [
    "assert img_tensor.shape == F.conv2d(img_tensor, vertical_filter, padding='same').shape\n",
    "assert img_tensor.shape == F.conv2d(img_tensor, vertical_filter, padding=1).shape\n",
    "assert not img_tensor.shape == F.conv2d(img_tensor, vertical_filter).shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "RNDcsnUDSo8P",
    "outputId": "d5679ad9-915b-4b33-f6fa-f3c46dd58fd7"
   },
   "outputs": [],
   "source": [
    "img_tensor.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "ebN_qqaRSW5W",
    "outputId": "7425bf47-ced8-457c-fef7-f9bb2c4afd62"
   },
   "outputs": [],
   "source": [
    "F.conv2d(img_tensor, vertical_filter, padding=1, stride=1).shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "d5t55mbrSNDM",
    "outputId": "56baeeed-ee03-43de-c008-95d12afbbc90"
   },
   "outputs": [],
   "source": [
    "F.conv2d(img_tensor, vertical_filter, padding=1, stride=2).shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/",
     "height": 517
    },
    "id": "JWEu1x09mkGh",
    "outputId": "095fa94c-6ccf-4a2c-93af-a2ba44f4c355"
   },
   "outputs": [],
   "source": [
    "url = 'https://www.syfy.com/sites/syfy/files/styles/scale--1200/public/cast_futurama_bender_0.jpg'\n",
    "response = requests.get(url)\n",
    "img = Image.open(BytesIO(response.content)).resize((350, 500))\n",
    "img"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "y6j7r7iuPbH8",
    "outputId": "58ab8e01-a312-4dd5-96ca-9c51935743b9"
   },
   "outputs": [],
   "source": [
    "img_tensor = to_tensor(img).unsqueeze(0)\n",
    "img_tensor.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "id": "uYr4l75QgWky"
   },
   "source": [
    "# Exercise 10.2\n",
    "\n",
    "Instead of explicitly defining filters, we'll just use random numbers this time.\n",
    "Based on the `F.conv2d` documentation, the filters should be of shape `(out_channels, in_channels, kernel/filter_height, kernel/filter_width)`.\n",
    "\n",
    "Complete the code in the cell below to create a filter with a 3x3 kernel that takes in a 3-channel image and returns a 16-channel feature map.\n",
    "We will use `F.conv2d` to generate the feature map.\n",
    "\n",
    "<!-- startquestion -->"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "SEJYUyPwhUdx",
    "outputId": "8669d025-4237-4337-a60c-54ead2fa97e1"
   },
   "outputs": [],
   "source": [
    "out_channels = ...\n",
    "in_channels = ...\n",
    "kH = ...\n",
    "kW = ...\n",
    "try:\n",
    "    random_filters = torch.rand((out_channels, in_channels, kH, kW))\n",
    "    print(random_filters.shape)\n",
    "except:\n",
    "    print('Please assign out_channels, in_channels, kH, and kW.')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "id": "hvHjxp5giXK7"
   },
   "outputs": [],
   "source": [
    "outputs = F.conv2d(img_tensor, random_filters, padding=1, stride=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "id": "fxI_1T0tPhQQ"
   },
   "outputs": [],
   "source": [
    "assert outputs.shape[1] == 16"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {
    "id": "UEOPp8lgQQwq"
   },
   "outputs": [],
   "source": [
    "conv_kwargs = {\n",
    "    'padding': 1,\n",
    "    'stride': 1\n",
    "}\n",
    "\n",
    "conv_layer = nn.Conv2d(3, 16, 3, **conv_kwargs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "DeR1-mO_2WwO",
    "outputId": "1ada9907-7579-4512-d7f8-c07e944b5f63"
   },
   "outputs": [],
   "source": [
    "conv_layer_output = conv_layer(img_tensor)\n",
    "conv_layer_output.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {
    "id": "PRvH0jGyqoW3"
   },
   "outputs": [],
   "source": [
    "conv_function_output = F.conv2d(img_tensor, conv_layer.weight.data, conv_layer.bias.data, **conv_kwargs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "id": "v7Q0PmTJqxQP"
   },
   "outputs": [],
   "source": [
    "assert (conv_function_output == conv_layer_output).all()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {
    "id": "fbOFVavcUxkS"
   },
   "outputs": [],
   "source": [
    "maxpool = nn.MaxPool2d(kernel_size=2, stride=2)\n",
    "avgpool = nn.AvgPool2d(kernel_size=2, stride=2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "7R_fH_seVFoC",
    "outputId": "1b58bf66-5be8-4722-f608-ffc8fd96397d"
   },
   "outputs": [],
   "source": [
    "print(f\"\"\"\n",
    "Oringinal image shape: {img_tensor.shape}\n",
    "Max pool tensor shape: {maxpool(img_tensor).shape}\n",
    "Avg pool tensor shape: {avgpool(img_tensor).shape}\n",
    "\"\"\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {
    "id": "krpaM9ufWG89"
   },
   "outputs": [],
   "source": [
    "conv_stride_2 = nn.Conv2d(3, 3, 3, 2, 1)\n",
    "conv_then_pool = nn.Sequential(\n",
    "    nn.Conv2d(3, 3, 3, 1, 1),\n",
    "    nn.MaxPool2d(2, 2)\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "rCWGU1JvW4C3",
    "outputId": "82c5b1e3-b367-4b5b-b60f-8a30e6cfc276"
   },
   "outputs": [],
   "source": [
    "print(f\"\"\"\n",
    "Shape of conv with a stride of 2: {conv_stride_2(img_tensor).shape}\n",
    "Shape of conv with a poolling layer: {conv_then_pool(img_tensor).shape}\n",
    "\"\"\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {
    "id": "uz5vlmPNfS6e"
   },
   "outputs": [],
   "source": [
    "adapt_avg_pool = nn.AdaptiveAvgPool2d(1)\n",
    "adapt_max_pool = nn.AdaptiveMaxPool2d(1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "U_IEGeE2f7N2",
    "outputId": "267ec43e-d963-4ce9-8bf8-6e2e9685d881"
   },
   "outputs": [],
   "source": [
    "outputs.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "K7Owh1gGfXiV",
    "outputId": "1182f1c4-f588-48e8-868d-188c042274f6"
   },
   "outputs": [],
   "source": [
    "adapt_avg_pool(outputs).shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {
    "colab": {
     "base_uri": "https://localhost:8080/"
    },
    "id": "FqKoPFFef5By",
    "outputId": "d20d0d9d-7360-44b7-e783-80782faf0d6f"
   },
   "outputs": [],
   "source": [
    "adapt_pool_with_flatten = nn.Sequential(\n",
    "    nn.AdaptiveMaxPool2d(1),\n",
    "    nn.Flatten()\n",
    ")\n",
    "\n",
    "print(adapt_pool_with_flatten(outputs).shape)"
   ]
  }
 ],
 "metadata": {
  "colab": {
   "authorship_tag": "ABX9TyOkcSPPpAfpY1mBRv9xonfp",
   "include_colab_link": true,
   "name": "10_Introduction_to_Convolutional_Neural_Nets.ipynb",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3",
   "name": "python3"
  },
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}