pytorch_mini_example.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": []
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "Example inspired by \n",
        "[PyTorch Tutorial](https://pytorch.org/tutorials/beginner/basics/quickstart_tutorial.html) for explaining ML terminology."
      ],
      "metadata": {
        "id": "F0QWRg6qo2Pp"
      }
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 36
        },
        "id": "huiFTKC9jZtF",
        "outputId": "d5f7f48d-2450-4b9c-cf76-ceaf0b804672"
      },
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "'2.0.0+cu118'"
            ],
            "application/vnd.google.colaboratory.intrinsic+json": {
              "type": "string"
            }
          },
          "metadata": {},
          "execution_count": 62
        }
      ],
      "source": [
        "import torch\n",
        "from torch import nn\n",
        "from torch.utils.data import DataLoader\n",
        "from torchvision import datasets\n",
        "from torchvision.transforms import ToTensor\n",
        "import numpy as np\n",
        "import matplotlib.pyplot as plt\n",
        "\n",
        "bg = \"#383838\"\n",
        "plt.rcParams.update({\n",
        "    \"figure.facecolor\": bg,\n",
        "    \"figure.edgecolor\": bg,\n",
        "    \"savefig.facecolor\": bg,\n",
        "    \"savefig.edgecolor\": bg})\n",
        "\n",
        "torch.__version__"
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "ds = datasets.MNIST\n",
        "# ds = datasets.FashionMNIST\n",
        "training_data = ds(\n",
        "    root=\"data\",\n",
        "    train=True,\n",
        "    download=True,\n",
        "    transform=ToTensor(),\n",
        ")\n",
        "test_data = ds(\n",
        "    root=\"data\",\n",
        "    train=False,\n",
        "    download=True,\n",
        "    transform=ToTensor(),\n",
        ")\n",
        "\n",
        "len(training_data), len(test_data)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "7MsNxxIrjkpS",
        "outputId": "ee111053-a7ce-4525-b323-c23484b10cc5"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(60000, 10000)"
            ]
          },
          "metadata": {},
          "execution_count": 63
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "first = training_data[0]\n",
        "first[0].shape, first[1]"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "wUyiAfE2pZya",
        "outputId": "38609d4b-50b7-483b-8951-6fb3584b56c3"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(torch.Size([1, 28, 28]), 5)"
            ]
          },
          "metadata": {},
          "execution_count": 64
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "for i in range(10):\n",
        "  print(training_data[i][1], end=',    ')\n",
        "\n",
        "  plt.subplot(1, 10, i+1)\n",
        "  plt.axis('off')\n",
        "  plt.imshow(training_data[i][0][0], cmap='gray')"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 97
        },
        "id": "YtzwG8SPj_w6",
        "outputId": "f60b2c2d-0478-4bba-b51e-89327b2036b8"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "5,    0,    4,    1,    9,    2,    1,    3,    1,    4,    "
          ]
        },
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 640x480 with 10 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {}
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "class NeuralNetwork(nn.Module):\n",
        "    def __init__(self):\n",
        "        super().__init__()\n",
        "        self.flatten = nn.Flatten()\n",
        "        self.layers = nn.Linear(28*28, 10) # for each pixel in image and each output category multiply it by weight (w), then for each category add bias (b)\n",
        "        # self.layers = nn.Sequential(\n",
        "        #     nn.Linear(28*28, 10),\n",
        "        #     nn.ReLU(),\n",
        "        #     # ...\n",
        "        # )\n",
        "\n",
        "    def forward(self, x):\n",
        "        x = self.flatten(x)\n",
        "        logits = self.layers(x)\n",
        "        return logits\n",
        "\n",
        "model = NeuralNetwork()\n",
        "model"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "vxqhFCzqoY-J",
        "outputId": "f676e7f7-2a24-47c4-99a1-fbf1b86eca47"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "NeuralNetwork(\n",
              "  (flatten): Flatten(start_dim=1, end_dim=-1)\n",
              "  (layers): Linear(in_features=784, out_features=10, bias=True)\n",
              ")"
            ]
          },
          "metadata": {},
          "execution_count": 66
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "# model.layers.weight.data.fill_(0.01)\n",
        "# model.layers.bias.data.fill_(0.01)\n",
        "model.layers.weight.data, model.layers.bias.data"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "oGIzvKsjrYC5",
        "outputId": "4fe8f572-652d-47fc-90cc-8e336764251b"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(tensor([[-9.0349e-03,  5.3291e-03,  2.5955e-02,  ..., -3.3422e-03,\n",
              "          -5.6225e-03,  1.6547e-02],\n",
              "         [ 3.4052e-02,  2.3017e-02, -1.4675e-02,  ...,  1.6260e-02,\n",
              "           3.2667e-02, -4.8776e-03],\n",
              "         [-1.3981e-02, -7.2632e-03, -1.6890e-02,  ..., -2.1339e-02,\n",
              "          -3.6693e-03,  2.2111e-02],\n",
              "         ...,\n",
              "         [ 2.1930e-02, -1.7316e-02, -2.8703e-03,  ..., -3.3779e-05,\n",
              "           3.3387e-02, -2.3545e-03],\n",
              "         [ 2.8816e-03,  1.7490e-02,  2.3678e-02,  ..., -1.8226e-02,\n",
              "          -3.2558e-02,  2.4380e-03],\n",
              "         [ 1.5006e-02, -2.6897e-02,  4.0560e-03,  ...,  2.6650e-02,\n",
              "           1.7819e-02, -1.1566e-02]]),\n",
              " tensor([-0.0274,  0.0078,  0.0356,  0.0257,  0.0106, -0.0273,  0.0267, -0.0308,\n",
              "         -0.0299, -0.0142]))"
            ]
          },
          "metadata": {},
          "execution_count": 67
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "x = training_data[0][0]\n",
        "y = torch.zeros(10)\n",
        "y[training_data[0][1]] = 1\n",
        "y # expected output as one-hot encoded probabilities"
      ],
      "metadata": {
        "id": "AsRmVNbqrgA2",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "1d23a478-69a1-466e-e40e-444e8052232b"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([0., 0., 0., 0., 0., 1., 0., 0., 0., 0.])"
            ]
          },
          "metadata": {},
          "execution_count": 68
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "out = model(x)\n",
        "list(enumerate(*out.detach().numpy())) # actual output from model => logits"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "P8IM5vU90ZtH",
        "outputId": "840d70e4-b9e4-4c70-f901-49e0427a0165"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[(0, 0.1771769),\n",
              " (1, -0.19832423),\n",
              " (2, -0.20448236),\n",
              " (3, 0.28688264),\n",
              " (4, 0.18993436),\n",
              " (5, 0.19386116),\n",
              " (6, 0.18575034),\n",
              " (7, -0.05542535),\n",
              " (8, -0.13201796),\n",
              " (9, -0.017122548)]"
            ]
          },
          "metadata": {},
          "execution_count": 69
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "torch.argmax(out) # index of max logit => predicted category"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "nr4WC7UcvbL6",
        "outputId": "8978bf52-6bc4-4002-b86b-c28da737e8e5"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor(3)"
            ]
          },
          "metadata": {},
          "execution_count": 70
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "torch.softmax(out, 1) # probabilities (convert logits by \"some magic\" so that all values are between 0 and 1, and the sum is 1)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "I9snAkkHweys",
        "outputId": "141c2985-c63a-4617-989b-d0ddf4552abb"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([[0.1127, 0.0774, 0.0769, 0.1258, 0.1141, 0.1146, 0.1137, 0.0893, 0.0827,\n",
              "         0.0928]], grad_fn=<SoftmaxBackward0>)"
            ]
          },
          "metadata": {},
          "execution_count": 71
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "-torch.log(torch.softmax(out, 1)) # negative logarithm of probabilities => entropy"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "EtdhzLagxIVU",
        "outputId": "5adbc1c9-115b-43a3-9f10-0e9d45cfb7f3"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([[2.1831, 2.5586, 2.5648, 2.0734, 2.1703, 2.1664, 2.1745, 2.4157, 2.4923,\n",
              "         2.3774]], grad_fn=<NegBackward0>)"
            ]
          },
          "metadata": {},
          "execution_count": 72
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "loss_fn = nn.CrossEntropyLoss() # a function that we want to minimize (for every model output on training data compared to desired output)\n",
        "\n",
        "optimizer = torch.optim.Adam(model.parameters(), lr=0.001) # side quest: explain stochastic gradient descent and learning rate\n",
        "\n",
        "train_dataloader = DataLoader(training_data, batch_size=100) # hyper-parametes used in training (\"search for program\") vs parameters inside the model (\"the program we found\")\n",
        "test_dataloader = DataLoader(test_data, batch_size=100)\n",
        "\n",
        "for epoch in range(1): # epochs vs batches\n",
        "    model.train()\n",
        "    for batch, (X, y) in enumerate(train_dataloader):\n",
        "        # X, y = X.to(device), y.to(device)  if we would be using GPU/MPU/TPU\n",
        "        pred = model(X)\n",
        "        loss = loss_fn(pred, y)\n",
        "        optimizer.zero_grad()\n",
        "        loss.backward()\n",
        "        optimizer.step()\n",
        "\n",
        "# model prediction after training\n",
        "out = model(x)\n",
        "list(enumerate(*out.detach().numpy()))"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "mJekqPFo2azh",
        "outputId": "abc63553-038e-441c-f198-088d709e6eb7"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[(0, -0.59068465),\n",
              " (1, -3.2879238),\n",
              " (2, -1.3059332),\n",
              " (3, 2.4671628),\n",
              " (4, -5.3817625),\n",
              " (5, 2.6767466),\n",
              " (6, -2.2072687),\n",
              " (7, -0.09491294),\n",
              " (8, -0.7049214),\n",
              " (9, -2.1103463)]"
            ]
          },
          "metadata": {},
          "execution_count": 73
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "# 7840 weights after training, visualized as 10 images for output neurons\n",
        "# red: excitation from a pixel in input image, blue: inhibition, black: no effect\n",
        "\n",
        "w = model.layers.weight.detach().numpy()\n",
        "nmin, nmax = w.min(), w.max()\n",
        "plt.figure(figsize=(28, 28))\n",
        "for i in range(10):\n",
        "  n = w[i].reshape(28, 28)\n",
        "  r = n * (n > 0) / nmax * 255\n",
        "  b = n * (n < 0) / nmin * 255\n",
        "  g = r / 4 + b / 2\n",
        "  img = np.uint8(np.dstack((r, g, b)))\n",
        "  plt.subplot(1, 10, i + 1)\n",
        "  plt.axis('off')\n",
        "  plt.imshow(img)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 172
        },
        "id": "38c03NKz8Wh0",
        "outputId": "1d476767-8839-440a-a68f-0f1921cf4185"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 2800x2800 with 10 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {}
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "model.layers.bias"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "4g8Uu0QEAMsr",
        "outputId": "4e1db780-7237-404e-a39a-ffa2539943d1"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "Parameter containing:\n",
              "tensor([-0.1371,  0.1987, -0.0054, -0.0668,  0.0719,  0.1001, -0.0092,  0.0777,\n",
              "        -0.2176, -0.0323], requires_grad=True)"
            ]
          },
          "metadata": {},
          "execution_count": 75
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "np.unique(np.array([i[1] for i in training_data]), return_counts=True) # frequency of each digit in the training data for comparison with bias"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "FRuvjJcA-SRV",
        "outputId": "d47b2359-d4e8-4417-d120-3b3ec20eb2b7"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "(array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]),\n",
              " array([5923, 6742, 5958, 6131, 5842, 5421, 5918, 6265, 5851, 5949]))"
            ]
          },
          "metadata": {},
          "execution_count": 76
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "# evaluation on test data\n",
        "\n",
        "model.eval()\n",
        "test_loss, correct = 0, 0\n",
        "with torch.no_grad():\n",
        "    for X, y in test_dataloader:\n",
        "        pred = model(X)\n",
        "        test_loss += loss_fn(pred, y).item()\n",
        "        correct += (pred.argmax(1) == y).type(torch.float).sum().item()\n",
        "test_loss /= len(test_dataloader)\n",
        "correct /= len(test_dataloader.dataset)\n",
        "print(f\"Test accuracy: {(100*correct):>0.1f}%, avg loss: {test_loss:>4f}\")"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "4tYmYmoI8dLa",
        "outputId": "bff02a2b-e0e7-4613-8e91-e7e98cf16467"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Test accuracy: 90.3%, avg loss: 0.365307\n"
          ]
        }
      ]
    }
  ]
}