examples/introduction.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Introduction to PyTorch and Poutyne\n",
    "\n",
    "In this notebook, we train a simple fully-connected network and a simple convolutional network on MNIST. First, we train it by coding our own training loop as the PyTorch library expects of us to. Then, we use Poutyne to simplify our code.\n",
    "\n",
    "Let's install the latest version of Poutyne (if it's not already) and import all the needed packages."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%pip install --upgrade poutyne\n",
    "# %pip install --upgrade colorama # Uncomment this line to enable colors.\n",
    "%matplotlib inline\n",
    "import os\n",
    "import math\n",
    "import shutil\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.optim as optim\n",
    "from torch.utils.data import random_split, DataLoader\n",
    "from torchvision import transforms, utils\n",
    "from torchvision.datasets.mnist import MNIST\n",
    "\n",
    "from poutyne import set_seeds, Model, ModelCheckpoint, CSVLogger, ModelBundle"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "Also, we need to set Python's, NumPy's and PyTorch's seeds by using Poutyne function so that our training is (almost) reproducible."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "set_seeds(42)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Basis of Training a Neural Network\n",
    "\n",
    "In **stochastic gradient descent**, a **batch** of `m` examples are drawn from the train dataset. In the so-called forward pass, these examples are passed through the neural network and an average of their loss values is done. In the backward pass, the average loss is backpropagated through the network to compute the gradient of each parameter. In practice, the `m` examples of a batch are drawn without replacement. Thus, we define one **epoch** of training being the number of batches needed to loop through the entire training dataset.\n",
    "\n",
    "In addition to the training dataset, a **validation dataset** is used to evaluate the neural network at the end of each epoch. This validation dataset can be used to select the best model during training and thus avoiding overfitting the training set. It also can have other uses such as selecting hyperparameters\n",
    "\n",
    "Finally, a **test dataset** is used at the end to evaluate the final model.\n",
    "\n",
    "### Training constants"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "pycharm": {
     "name": "#%% md\n"
    }
   },
   "source": [
    "Now, let's set our training constants. We first have the CUDA device used for training if one is present. Second, we set the train_split to 0.8 (80%) to use 80% of the dataset for training and 20% for testing the trained model. Third, we set the number of classes (i.e. one for each number). Finally, we set the batch size (i.e. the number of elements to see before updating the model), the learning rate for the optimizer, and the number of epochs (i.e. the number of times we see the full dataset).\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "cuda_device = 0\n",
    "device = torch.device(\"cuda:%d\" % cuda_device if torch.cuda.is_available() else \"cpu\")\n",
    "\n",
    "train_split_percent = 0.8\n",
    "\n",
    "num_classes = 10\n",
    "\n",
    "batch_size = 32\n",
    "learning_rate = 0.1\n",
    "num_epochs = 5"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Loading the MNIST dataset\n",
    "\n",
    "The following code helps load the MNIST dataset and creates the PyTorch DataLoaders that split our datasets into batches. The train DataLoader shuffles the examples of the train dataset to draw the examples without replacement."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "full_train_dataset = MNIST('./datasets/', train=True, download=True, transform=transforms.ToTensor())\n",
    "test_dataset = MNIST('./datasets/', train=False, download=True, transform=transforms.ToTensor())\n",
    "\n",
    "num_data = len(full_train_dataset)\n",
    "train_length = int(math.floor(train_split_percent * num_data))\n",
    "valid_length = num_data - train_length\n",
    "\n",
    "train_dataset, valid_dataset = random_split(\n",
    "    full_train_dataset,\n",
    "    [train_length, valid_length],\n",
    "    generator=torch.Generator().manual_seed(42),\n",
    ")\n",
    "\n",
    "train_loader = DataLoader(train_dataset, batch_size=batch_size, num_workers=2, shuffle=True)\n",
    "valid_loader = DataLoader(valid_dataset, batch_size=batch_size, num_workers=2)\n",
    "test_loader = DataLoader(test_dataset, batch_size=batch_size, num_workers=2)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's look at some examples of the dataset by looking at the first batch in our train DataLoader and formatting it into a grid and plotting it."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x720 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "inputs = next(iter(train_loader))[0]\n",
    "input_grid = utils.make_grid(inputs)\n",
    "\n",
    "fig = plt.figure(figsize=(10, 10))\n",
    "inp = input_grid.numpy().transpose((1, 2, 0))\n",
    "plt.imshow(inp)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Neural Network Architectures\n",
    "\n",
    "We train a fully-connected neural network and a convolutional neural network with approximately the same number of parameters.\n",
    "\n",
    "#### Fully-connected Network\n",
    "In short, the fully-connected network follows this architecture: ``Input -> [Linear -> ReLU]*3 -> Linear``. The following table shows it in details:\n",
    "\n",
    "| Layer Type                  | Output Size |       # of Parameters       |\n",
    "|-----------------------------|:-----------:|:---------------------------:|\n",
    "| Input                       |   1x28x28   |              0              |\n",
    "| Flatten                     |  1\\*28\\*28  |              0              |\n",
    "| **Linear with 256 neurons** |     256     | 28\\*28\\*256 + 256 = 200,960 |\n",
    "| ReLU                        |      *      |              0              |\n",
    "| **Linear with 128 neurons** |     128     |   256*128 + 128 = 32,896    |\n",
    "| ReLU                        |      *      |              0              |\n",
    "| **Linear with 64 neurons**  |     64      |     128*64 + 64 = 8,256     |\n",
    "| ReLU                        |      *      |              0              |\n",
    "| **Linear with 10 neurons**  |      10     |      64*10 + 10 = 650       |\n",
    "\n",
    "Total # of parameters of the fully-connected network: 242,762\n",
    "\n",
    "#### Convolutional Network\n",
    "\n",
    "The convolutional neural network architecture starts with some convolution and max-pooling layers. These are then followed by fully-connected layers. We calculate the total number of parameters that the network needs. In short, the convolutional network follows this architecture: ``Input -> [Conv -> ReLU -> MaxPool]*2 -> Dropout -> Linear -> ReLU -> Dropout -> Linear``. The following table shows it in details:\n",
    "\n",
    "| Layer Type                                     | Output Size |        # of Parameters        |\n",
    "|------------------------------------------------|:-----------:|:-----------------------------:|\n",
    "| Input                                          |   1x28x28   |               0               |\n",
    "| **Conv with 16 3x3 filters with padding of 1** |   16x28x28  |    16\\*1\\*3\\*3 + 16 = 160     |\n",
    "| ReLU                                           |   16x28x28  |               0               |\n",
    "| MaxPool 2x2                                    |   16x14x14  |               0               |\n",
    "| **Conv with 32 3x3 filters with padding of 1** |   32x14x14  |   32\\*16\\*3\\*3 + 32 = 4,640   |\n",
    "| ReLU                                           |   32x14x14  |               0               |\n",
    "| MaxPool 2x2                                    |    32x7x7   |               0               |\n",
    "| Dropout of 0.25                                |    32x7x7   |               0               |\n",
    "| Flatten                                        |   32\\*7\\*7  |               0               |\n",
    "| **Linear with 128 neurons**                    |     128     | 32\\*7\\*7\\*128 + 128 = 200,832 |\n",
    "| ReLU                                           |     128     |               0               |\n",
    "| Dropout of 0.5                                 |     128     |               0               |\n",
    "| **Linear with 10 neurons**                     |      10     |        128\\*10 + 10 = 1290    |\n",
    "\n",
    "Total # of parameters of the convolutional network: 206,922"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "def create_fully_connected_network():\n",
    "    \"\"\"\n",
    "    This function returns the fully-connected network layed out above.\n",
    "    \"\"\"\n",
    "    return nn.Sequential(\n",
    "        nn.Flatten(),\n",
    "        nn.Linear(28 * 28, 256),\n",
    "        nn.ReLU(),\n",
    "        nn.Linear(256, 128),\n",
    "        nn.ReLU(),\n",
    "        nn.Linear(128, 64),\n",
    "        nn.ReLU(),\n",
    "        nn.Linear(64, num_classes),\n",
    "    )\n",
    "\n",
    "\n",
    "def create_convolutional_network():\n",
    "    \"\"\"\n",
    "    This function returns the convolutional network layed out above.\n",
    "    \"\"\"\n",
    "    return nn.Sequential(\n",
    "        nn.Conv2d(in_channels=1, out_channels=16, kernel_size=3, padding=1),\n",
    "        nn.ReLU(),\n",
    "        nn.MaxPool2d(2),\n",
    "        nn.Conv2d(in_channels=16, out_channels=32, kernel_size=3, padding=1),\n",
    "        nn.ReLU(),\n",
    "        nn.MaxPool2d(2),\n",
    "        nn.Dropout(0.25),\n",
    "        nn.Flatten(),\n",
    "        nn.Linear(32 * 7 * 7, 128),\n",
    "        nn.ReLU(),\n",
    "        nn.Dropout(0.5),\n",
    "        nn.Linear(128, num_classes),\n",
    "    )"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training the PyTorch way \n",
    "\n",
    "That is, writing your own training loop."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "def pytorch_accuracy(y_pred, y_true):\n",
    "    \"\"\"\n",
    "    Computes the accuracy for a batch of predictions\n",
    "\n",
    "    Args:\n",
    "        y_pred (torch.Tensor): the logit predictions of the neural network.\n",
    "        y_true (torch.Tensor): the ground truths.\n",
    "\n",
    "    Returns:\n",
    "        The average accuracy of the batch.\n",
    "    \"\"\"\n",
    "    y_pred = y_pred.argmax(1)\n",
    "    return (y_pred == y_true).float().mean() * 100\n",
    "\n",
    "\n",
    "def pytorch_train_one_epoch(network, optimizer, loss_function):\n",
    "    \"\"\"\n",
    "    Trains the neural network for one epoch on the train DataLoader.\n",
    "\n",
    "    Args:\n",
    "        network (torch.nn.Module): The neural network to train.\n",
    "        optimizer (torch.optim.Optimizer): The optimizer of the neural network\n",
    "        loss_function: The loss function.\n",
    "\n",
    "    Returns:\n",
    "        A tuple (loss, accuracy) corresponding to an average of the losses and\n",
    "        an average of the accuracy, respectively, on the train DataLoader.\n",
    "    \"\"\"\n",
    "    network.train(True)\n",
    "    with torch.enable_grad():\n",
    "        loss_sum = 0.0\n",
    "        acc_sum = 0.0\n",
    "        example_count = 0\n",
    "        for x, y in train_loader:\n",
    "            # Transfer batch on GPU if needed.\n",
    "            x = x.to(device)\n",
    "            y = y.to(device)\n",
    "\n",
    "            optimizer.zero_grad()\n",
    "\n",
    "            y_pred = network(x)\n",
    "\n",
    "            loss = loss_function(y_pred, y)\n",
    "\n",
    "            loss.backward()\n",
    "\n",
    "            optimizer.step()\n",
    "\n",
    "            # Since the loss and accuracy are averages for the batch, we multiply\n",
    "            # them by the the number of examples so that we can do the right\n",
    "            # averages at the end of the epoch.\n",
    "            loss_sum += float(loss) * len(x)\n",
    "            acc_sum += float(pytorch_accuracy(y_pred, y)) * len(x)\n",
    "            example_count += len(x)\n",
    "\n",
    "    avg_loss = loss_sum / example_count\n",
    "    avg_acc = acc_sum / example_count\n",
    "    return avg_loss, avg_acc\n",
    "\n",
    "\n",
    "def pytorch_test(network, loader, loss_function):\n",
    "    \"\"\"\n",
    "    Tests the neural network on a DataLoader.\n",
    "\n",
    "    Args:\n",
    "        network (torch.nn.Module): The neural network to test.\n",
    "        loader (torch.utils.data.DataLoader): The DataLoader to test on.\n",
    "        loss_function: The loss function.\n",
    "\n",
    "    Returns:\n",
    "        A tuple (loss, accuracy) corresponding to an average of the losses and\n",
    "        an average of the accuracy, respectively, on the DataLoader.\n",
    "    \"\"\"\n",
    "    network.eval()\n",
    "    with torch.no_grad():\n",
    "        loss_sum = 0.0\n",
    "        acc_sum = 0.0\n",
    "        example_count = 0\n",
    "        for x, y in loader:\n",
    "            # Transfer batch on GPU if needed.\n",
    "            x = x.to(device)\n",
    "            y = y.to(device)\n",
    "\n",
    "            y_pred = network(x)\n",
    "            loss = loss_function(y_pred, y)\n",
    "\n",
    "            # Since the loss and accuracy are averages for the batch, we multiply\n",
    "            # them by the the number of examples so that we can do the right\n",
    "            # averages at the end of the test.\n",
    "            loss_sum += float(loss) * len(x)\n",
    "            acc_sum += float(pytorch_accuracy(y_pred, y)) * len(x)\n",
    "            example_count += len(x)\n",
    "\n",
    "    avg_loss = loss_sum / example_count\n",
    "    avg_acc = acc_sum / example_count\n",
    "    return avg_loss, avg_acc\n",
    "\n",
    "\n",
    "def pytorch_train(network):\n",
    "    \"\"\"\n",
    "    This function transfers the neural network to the right device,\n",
    "    trains it for a certain number of epochs, tests at each epoch on\n",
    "    the validation set and outputs the results on the test set at the\n",
    "    end of training.\n",
    "\n",
    "    Args:\n",
    "        network (torch.nn.Module): The neural network to train.\n",
    "\n",
    "    Example:\n",
    "        This function displays something like this:\n",
    "\n",
    "        .. code-block:: python\n",
    "\n",
    "            Epoch 1/5: loss: 0.5026924496193726, acc: 84.26666259765625, val_loss: 0.17258917854229608, val_acc: 94.75\n",
    "            Epoch 2/5: loss: 0.13690324830015502, acc: 95.73332977294922, val_loss: 0.14024296019474666, val_acc: 95.68333435058594\n",
    "            Epoch 3/5: loss: 0.08836929737279813, acc: 97.29582977294922, val_loss: 0.10380942322810491, val_acc: 96.66666412353516\n",
    "            Epoch 4/5: loss: 0.06714504160980383, acc: 97.91874694824219, val_loss: 0.09626663728555043, val_acc: 97.18333435058594\n",
    "            Epoch 5/5: loss: 0.05063822727650404, acc: 98.42708587646484, val_loss: 0.10017542181412378, val_acc: 96.95833587646484\n",
    "            Test:\n",
    "                Loss: 0.09501855444908142\n",
    "                Accuracy: 97.12999725341797\n",
    "    \"\"\"\n",
    "    print(network)\n",
    "\n",
    "    # Transfer weights on GPU if needed.\n",
    "    network.to(device)\n",
    "\n",
    "    optimizer = optim.SGD(network.parameters(), lr=learning_rate)\n",
    "    loss_function = nn.CrossEntropyLoss()\n",
    "\n",
    "    for epoch in range(1, num_epochs + 1):\n",
    "        # Training the neural network via backpropagation\n",
    "        train_loss, train_acc = pytorch_train_one_epoch(network, optimizer, loss_function)\n",
    "\n",
    "        # Validation at the end of the epoch\n",
    "        valid_loss, valid_acc = pytorch_test(network, valid_loader, loss_function)\n",
    "\n",
    "        print(\n",
    "            \"Epoch {}/{}: loss: {}, acc: {}, val_loss: {}, val_acc: {}\".format(\n",
    "                epoch, num_epochs, train_loss, train_acc, valid_loss, valid_acc\n",
    "            )\n",
    "        )\n",
    "\n",
    "    # Test at the end of the training\n",
    "    test_loss, test_acc = pytorch_test(network, test_loader, loss_function)\n",
    "    print('Test:\\n\\tLoss: {}\\n\\tAccuracy: {}'.format(test_loss, test_acc))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sequential(\n",
      "  (0): Flatten(start_dim=1, end_dim=-1)\n",
      "  (1): Linear(in_features=784, out_features=256, bias=True)\n",
      "  (2): ReLU()\n",
      "  (3): Linear(in_features=256, out_features=128, bias=True)\n",
      "  (4): ReLU()\n",
      "  (5): Linear(in_features=128, out_features=64, bias=True)\n",
      "  (6): ReLU()\n",
      "  (7): Linear(in_features=64, out_features=10, bias=True)\n",
      ")\n",
      "Epoch 1/5: loss: 0.5130350270022949, acc: 83.85416666666667, val_loss: 0.18178541194895903, val_acc: 94.46666666666667\n",
      "Epoch 2/5: loss: 0.1374620988739965, acc: 95.78541666666666, val_loss: 0.12427739042043685, val_acc: 96.225\n",
      "Epoch 3/5: loss: 0.09078938368932965, acc: 97.16666666666667, val_loss: 0.11353350764078399, val_acc: 96.56666666666666\n",
      "Epoch 4/5: loss: 0.0665546608583536, acc: 97.92916666666666, val_loss: 0.10379793403080353, val_acc: 96.925\n",
      "Epoch 5/5: loss: 0.05080128996776572, acc: 98.41041666666666, val_loss: 0.09510265534456509, val_acc: 97.25\n",
      "Test:\n",
      "\tLoss: 0.07824153664603364\n",
      "\tAccuracy: 97.67\n"
     ]
    }
   ],
   "source": [
    "fc_net = create_fully_connected_network()\n",
    "pytorch_train(fc_net)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sequential(\n",
      "  (0): Conv2d(1, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (1): ReLU()\n",
      "  (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (3): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (4): ReLU()\n",
      "  (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (6): Dropout(p=0.25, inplace=False)\n",
      "  (7): Flatten(start_dim=1, end_dim=-1)\n",
      "  (8): Linear(in_features=1568, out_features=128, bias=True)\n",
      "  (9): ReLU()\n",
      "  (10): Dropout(p=0.5, inplace=False)\n",
      "  (11): Linear(in_features=128, out_features=10, bias=True)\n",
      ")\n",
      "Epoch 1/5: loss: 0.36489138784228514, acc: 88.09583333333333, val_loss: 0.08525923740367095, val_acc: 97.49166666666666\n",
      "Epoch 2/5: loss: 0.13036188714050997, acc: 96.16458333333334, val_loss: 0.06550583391217515, val_acc: 98.10833333333333\n",
      "Epoch 3/5: loss: 0.10261346009676345, acc: 96.90625, val_loss: 0.05349772878705213, val_acc: 98.46666666666667\n",
      "Epoch 4/5: loss: 0.08726432745321654, acc: 97.39375, val_loss: 0.04868311001577725, val_acc: 98.56666666666666\n",
      "Epoch 5/5: loss: 0.07482324100522479, acc: 97.74166666666666, val_loss: 0.04666790096772214, val_acc: 98.69166666666666\n",
      "Test:\n",
      "\tLoss: 0.03658340354624088\n",
      "\tAccuracy: 98.84\n"
     ]
    }
   ],
   "source": [
    "conv_net = create_convolutional_network()\n",
    "pytorch_train(conv_net)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training the Poutyne way\n",
    "\n",
    "That is, only few lines of code with a better output."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "def poutyne_train(network):\n",
    "    \"\"\"\n",
    "    This function creates a Poutyne Model (see https://poutyne.org/model.html), sends the Model\n",
    "    on the specified device, and uses the `fit_generator` method to train the neural network.\n",
    "    At the end, the `evaluate_generator` is used on  the test set.\n",
    "\n",
    "    Args:\n",
    "        network (torch.nn.Module): The neural network to train.\n",
    "    \"\"\"\n",
    "    print(network)\n",
    "\n",
    "    optimizer = optim.SGD(network.parameters(), lr=learning_rate)\n",
    "    loss_function = nn.CrossEntropyLoss()\n",
    "\n",
    "    # Poutyne Model on GPU\n",
    "    model = Model(\n",
    "        network,\n",
    "        optimizer,\n",
    "        loss_function,\n",
    "        batch_metrics=['accuracy'],\n",
    "        device=device,\n",
    "    )\n",
    "\n",
    "    # Train\n",
    "    model.fit_generator(train_loader, valid_loader, epochs=num_epochs)\n",
    "\n",
    "    # Test\n",
    "    test_loss, test_acc = model.evaluate_generator(test_loader)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sequential(\n",
      "  (0): Flatten(start_dim=1, end_dim=-1)\n",
      "  (1): Linear(in_features=784, out_features=256, bias=True)\n",
      "  (2): ReLU()\n",
      "  (3): Linear(in_features=256, out_features=128, bias=True)\n",
      "  (4): ReLU()\n",
      "  (5): Linear(in_features=128, out_features=64, bias=True)\n",
      "  (6): ReLU()\n",
      "  (7): Linear(in_features=64, out_features=10, bias=True)\n",
      ")\n",
      "\u001b[35mEpoch: \u001b[36m1/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.20s \u001b[35mloss:\u001b[94m 0.483920\u001b[35m acc:\u001b[94m 84.966667\u001b[35m val_loss:\u001b[94m 0.196932\u001b[35m val_acc:\u001b[94m 93.983333\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m2/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.28s \u001b[35mloss:\u001b[94m 0.136902\u001b[35m acc:\u001b[94m 95.850000\u001b[35m val_loss:\u001b[94m 0.130139\u001b[35m val_acc:\u001b[94m 95.991667\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m3/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.17s \u001b[35mloss:\u001b[94m 0.088676\u001b[35m acc:\u001b[94m 97.231250\u001b[35m val_loss:\u001b[94m 0.109163\u001b[35m val_acc:\u001b[94m 96.533333\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m4/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.33s \u001b[35mloss:\u001b[94m 0.066252\u001b[35m acc:\u001b[94m 97.933333\u001b[35m val_loss:\u001b[94m 0.102234\u001b[35m val_acc:\u001b[94m 96.966667\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m5/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.23s \u001b[35mloss:\u001b[94m 0.049697\u001b[35m acc:\u001b[94m 98.439583\u001b[35m val_loss:\u001b[94m 0.091193\u001b[35m val_acc:\u001b[94m 97.225000\u001b[0m\n",
      "\u001b[35mTest steps: \u001b[36m313 \u001b[32m0.77s \u001b[35mtest_loss:\u001b[94m 0.081475\u001b[35m test_acc:\u001b[94m 97.610000\u001b[0m                                                   \n"
     ]
    }
   ],
   "source": [
    "fc_net = create_fully_connected_network()\n",
    "poutyne_train(fc_net)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "pycharm": {
     "name": "#%%\n"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sequential(\n",
      "  (0): Conv2d(1, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (1): ReLU()\n",
      "  (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (3): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (4): ReLU()\n",
      "  (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (6): Dropout(p=0.25, inplace=False)\n",
      "  (7): Flatten(start_dim=1, end_dim=-1)\n",
      "  (8): Linear(in_features=1568, out_features=128, bias=True)\n",
      "  (9): ReLU()\n",
      "  (10): Dropout(p=0.5, inplace=False)\n",
      "  (11): Linear(in_features=128, out_features=10, bias=True)\n",
      ")\n",
      "\u001b[35mEpoch: \u001b[36m1/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.43s \u001b[35mloss:\u001b[94m 0.369546\u001b[35m acc:\u001b[94m 88.070833\u001b[35m val_loss:\u001b[94m 0.118247\u001b[35m val_acc:\u001b[94m 96.400000\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m2/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.77s \u001b[35mloss:\u001b[94m 0.131069\u001b[35m acc:\u001b[94m 96.060417\u001b[35m val_loss:\u001b[94m 0.070504\u001b[35m val_acc:\u001b[94m 97.883333\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m3/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.46s \u001b[35mloss:\u001b[94m 0.102619\u001b[35m acc:\u001b[94m 96.785417\u001b[35m val_loss:\u001b[94m 0.056284\u001b[35m val_acc:\u001b[94m 98.458333\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m4/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.75s \u001b[35mloss:\u001b[94m 0.084918\u001b[35m acc:\u001b[94m 97.431250\u001b[35m val_loss:\u001b[94m 0.066149\u001b[35m val_acc:\u001b[94m 97.916667\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m5/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.49s \u001b[35mloss:\u001b[94m 0.076102\u001b[35m acc:\u001b[94m 97.731250\u001b[35m val_loss:\u001b[94m 0.046076\u001b[35m val_acc:\u001b[94m 98.591667\u001b[0m\n",
      "\u001b[35mTest steps: \u001b[36m313 \u001b[32m0.83s \u001b[35mtest_loss:\u001b[94m 0.032175\u001b[35m test_acc:\u001b[94m 98.920000\u001b[0m                                                  \n"
     ]
    }
   ],
   "source": [
    "conv_net = create_convolutional_network()\n",
    "poutyne_train(conv_net)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Poutyne Callbacks\n",
    "\n",
    "One nice feature of Poutyne is [callbacks](https://poutyne.org/callbacks.html). Callbacks allow to do actions during training of the neural network. In the following example, we use 3 callbacks. One that saves the latest weights in a file to be able to continue the optimization at the end of training if more epochs are needed. Another one that saves the best weights according to the performance on the validation dataset. Finally, another one that saves the displayed logs into a TSV file."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train_with_callbacks(name, network):\n",
    "    \"\"\"\n",
    "    In addition to the the `poutyne_train`, this function saves checkpoints and logs as described above.\n",
    "\n",
    "    Args:\n",
    "        name (str): a name used to save logs and checkpoints.\n",
    "        network (torch.nn.Module): The neural network to train.\n",
    "    \"\"\"\n",
    "    print(network)\n",
    "\n",
    "    # We are saving everything into ./saves/{name}.\n",
    "    save_path = os.path.join('saves', name)\n",
    "\n",
    "    # Creating saving directory if necessary.\n",
    "    os.makedirs(save_path, exist_ok=True)\n",
    "\n",
    "    callbacks = [\n",
    "        # Save the latest weights to be able to continue the optimization at the end for more epochs.\n",
    "        ModelCheckpoint(os.path.join(save_path, 'last_epoch.ckpt')),\n",
    "        # Save the weights in a new file when the current model is better than all previous models.\n",
    "        ModelCheckpoint(\n",
    "            os.path.join(save_path, 'best_epoch_{epoch}.ckpt'),\n",
    "            monitor='val_acc',\n",
    "            mode='max',\n",
    "            save_best_only=True,\n",
    "            restore_best=True,\n",
    "            verbose=True,\n",
    "        ),\n",
    "        # Save the losses and accuracies for each epoch in a TSV.\n",
    "        CSVLogger(os.path.join(save_path, 'log.tsv'), separator='\\t'),\n",
    "    ]\n",
    "\n",
    "    optimizer = optim.SGD(network.parameters(), lr=learning_rate)\n",
    "    loss_function = nn.CrossEntropyLoss()\n",
    "\n",
    "    model = Model(\n",
    "        network,\n",
    "        optimizer,\n",
    "        loss_function,\n",
    "        batch_metrics=['accuracy'],\n",
    "        device=device,\n",
    "    )\n",
    "    model.fit_generator(train_loader, valid_loader, epochs=num_epochs, callbacks=callbacks)\n",
    "\n",
    "    test_loss, test_acc = model.evaluate_generator(test_loader)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sequential(\n",
      "  (0): Flatten(start_dim=1, end_dim=-1)\n",
      "  (1): Linear(in_features=784, out_features=256, bias=True)\n",
      "  (2): ReLU()\n",
      "  (3): Linear(in_features=256, out_features=128, bias=True)\n",
      "  (4): ReLU()\n",
      "  (5): Linear(in_features=128, out_features=64, bias=True)\n",
      "  (6): ReLU()\n",
      "  (7): Linear(in_features=64, out_features=10, bias=True)\n",
      ")\n",
      "\u001b[35mEpoch: \u001b[36m1/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.32s \u001b[35mloss:\u001b[94m 0.501808\u001b[35m acc:\u001b[94m 84.033333\u001b[35m val_loss:\u001b[94m 0.194768\u001b[35m val_acc:\u001b[94m 93.983333\u001b[0m\n",
      "Epoch 1: val_acc improved from -inf to 93.98333, saving file to saves/fc/best_epoch_1.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m2/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.44s \u001b[35mloss:\u001b[94m 0.137684\u001b[35m acc:\u001b[94m 95.852083\u001b[35m val_loss:\u001b[94m 0.144413\u001b[35m val_acc:\u001b[94m 95.908333\u001b[0m\n",
      "Epoch 2: val_acc improved from 93.98333 to 95.90833, saving file to saves/fc/best_epoch_2.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m3/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.29s \u001b[35mloss:\u001b[94m 0.092897\u001b[35m acc:\u001b[94m 97.127083\u001b[35m val_loss:\u001b[94m 0.102404\u001b[35m val_acc:\u001b[94m 96.925000\u001b[0m\n",
      "Epoch 3: val_acc improved from 95.90833 to 96.92500, saving file to saves/fc/best_epoch_3.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m4/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.27s \u001b[35mloss:\u001b[94m 0.067247\u001b[35m acc:\u001b[94m 97.922917\u001b[35m val_loss:\u001b[94m 0.107454\u001b[35m val_acc:\u001b[94m 96.741667\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m5/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m5.19s \u001b[35mloss:\u001b[94m 0.049116\u001b[35m acc:\u001b[94m 98.416667\u001b[35m val_loss:\u001b[94m 0.094097\u001b[35m val_acc:\u001b[94m 97.175000\u001b[0m\n",
      "Epoch 5: val_acc improved from 96.92500 to 97.17500, saving file to saves/fc/best_epoch_5.ckpt\n",
      "Restoring data from saves/fc/best_epoch_5.ckpt\n",
      "\u001b[35mTest steps: \u001b[36m313 \u001b[32m0.77s \u001b[35mtest_loss:\u001b[94m 0.081783\u001b[35m test_acc:\u001b[94m 97.630000\u001b[0m                                                   \n"
     ]
    }
   ],
   "source": [
    "fc_net = create_fully_connected_network()\n",
    "train_with_callbacks('fc', fc_net)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sequential(\n",
      "  (0): Conv2d(1, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (1): ReLU()\n",
      "  (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (3): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (4): ReLU()\n",
      "  (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (6): Dropout(p=0.25, inplace=False)\n",
      "  (7): Flatten(start_dim=1, end_dim=-1)\n",
      "  (8): Linear(in_features=1568, out_features=128, bias=True)\n",
      "  (9): ReLU()\n",
      "  (10): Dropout(p=0.5, inplace=False)\n",
      "  (11): Linear(in_features=128, out_features=10, bias=True)\n",
      ")\n",
      "\u001b[35mEpoch: \u001b[36m1/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.81s \u001b[35mloss:\u001b[94m 0.366644\u001b[35m acc:\u001b[94m 88.206250\u001b[35m val_loss:\u001b[94m 0.099492\u001b[35m val_acc:\u001b[94m 96.816667\u001b[0m\n",
      "Epoch 1: val_acc improved from -inf to 96.81667, saving file to saves/conv/best_epoch_1.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m2/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.69s \u001b[35mloss:\u001b[94m 0.134940\u001b[35m acc:\u001b[94m 95.912500\u001b[35m val_loss:\u001b[94m 0.065306\u001b[35m val_acc:\u001b[94m 98.050000\u001b[0m\n",
      "Epoch 2: val_acc improved from 96.81667 to 98.05000, saving file to saves/conv/best_epoch_2.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m3/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.57s \u001b[35mloss:\u001b[94m 0.101323\u001b[35m acc:\u001b[94m 96.956250\u001b[35m val_loss:\u001b[94m 0.056368\u001b[35m val_acc:\u001b[94m 98.366667\u001b[0m\n",
      "Epoch 3: val_acc improved from 98.05000 to 98.36667, saving file to saves/conv/best_epoch_3.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m4/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.54s \u001b[35mloss:\u001b[94m 0.084502\u001b[35m acc:\u001b[94m 97.331250\u001b[35m val_loss:\u001b[94m 0.053686\u001b[35m val_acc:\u001b[94m 98.341667\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m5/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m6.71s \u001b[35mloss:\u001b[94m 0.076835\u001b[35m acc:\u001b[94m 97.654167\u001b[35m val_loss:\u001b[94m 0.045052\u001b[35m val_acc:\u001b[94m 98.666667\u001b[0m\n",
      "Epoch 5: val_acc improved from 98.36667 to 98.66667, saving file to saves/conv/best_epoch_5.ckpt\n",
      "Restoring data from saves/conv/best_epoch_5.ckpt\n",
      "\u001b[35mTest steps: \u001b[36m313 \u001b[32m0.82s \u001b[35mtest_loss:\u001b[94m 0.034237\u001b[35m test_acc:\u001b[94m 98.950000\u001b[0m                                                   \n"
     ]
    }
   ],
   "source": [
    "conv_net = create_convolutional_network()\n",
    "train_with_callbacks('conv', conv_net)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Making Your Own Callback\n",
    "\n",
    "While Poutyne provides a great number of [predefined callbacks](https://poutyne.org/callbacks.html), it is sometimes useful to make your own callback. In addition to the [documentation](https://poutyne.org/callbacks.html#poutyne.framework.callbacks.Callback) of the Callback class, see the [Tips and Tricks notebook](https://github.com/GRAAL-Research/poutyne/blob/master/examples/tips_and_tricks.ipynb) for an example."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Poutyne ModelBundle\n",
    "\n",
    "Most of the time when using Poutyne (or even Pytorch in general), we will find ourselves in an iterative model hyperparameters finetuning loop. For efficient model search, we will usually wish to save our best performing models, their training and testing statistics and even sometimes wish to retrain an already trained model for further tuning. All of the above can be easily implemented with the flexibility of Poutyne Callbacks, but having to define and initialize each and every Callback object we wish for our model quickly feels cumbersome.\n",
    "\n",
    "This is why Poutyne provides a [ModelBundle class](https://poutyne.org/experiment.html#poutyne.ModelBundle), which aims specifically at enabling quick model iteration search, while not sacrifying on the quality of a single experiment - statistics logging, best models saving, etc. ModelBundle is actually a simple wrapper between a PyTorch network and Poutyne's core Callback objects for logging and saving. Given a working directory where to output the various logging files and a PyTorch network, the ModelBundle class reduces the whole training loop to a single line.\n",
    "\n",
    "The following code uses [Poutyne's ModelBundle class](https://poutyne.org/experiment.html#poutyne.ModelBundle) to train a network for 5 epochs. The code is quite simpler than the code in the Poutyne Callbacks section while doing more (only 3 lines). Once trained for 5 epochs, it is then possible to resume the optimization at the 5th epoch for 5 more epochs until the 10th epoch using the same function."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train_model_bundle(network, name, epochs=5):\n",
    "    \"\"\"\n",
    "    This function creates a Poutyne ModelBundle, trains the input module\n",
    "    on the train loader and then tests its performance on the test loader.\n",
    "    All training and testing statistics are saved, as well as best model\n",
    "    checkpoints.\n",
    "\n",
    "    Args:\n",
    "        network (torch.nn.Module): The neural network to train.\n",
    "        working_directory (str): The directory where to output files to save.\n",
    "        epochs (int): The number of epochs. (Default: 5)\n",
    "    \"\"\"\n",
    "    print(network)\n",
    "\n",
    "    optimizer = optim.SGD(network.parameters(), lr=learning_rate)\n",
    "\n",
    "    # Everything is going to be saved in ./saves/{name}.\n",
    "    save_path = os.path.join('saves', name)\n",
    "\n",
    "    # Poutyne ModelBundle\n",
    "    model_bundle = ModelBundle.from_network(\n",
    "        save_path,\n",
    "        network,\n",
    "        device=device,\n",
    "        optimizer=optimizer,\n",
    "        task='classif',\n",
    "    )\n",
    "\n",
    "    # Train\n",
    "    model_bundle.train(train_loader, valid_loader, epochs=epochs)\n",
    "\n",
    "    # Test\n",
    "    model_bundle.test(test_loader)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's first remove the output directory if it exists so that we start training from scratch."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "shutil.rmtree('saves/conv_net_model_bundle', ignore_errors=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Let's train for 5 epochs."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sequential(\n",
      "  (0): Conv2d(1, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (1): ReLU()\n",
      "  (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (3): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (4): ReLU()\n",
      "  (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (6): Dropout(p=0.25, inplace=False)\n",
      "  (7): Flatten(start_dim=1, end_dim=-1)\n",
      "  (8): Linear(in_features=1568, out_features=128, bias=True)\n",
      "  (9): ReLU()\n",
      "  (10): Dropout(p=0.5, inplace=False)\n",
      "  (11): Linear(in_features=128, out_features=10, bias=True)\n",
      ")\n",
      "\u001b[35mEpoch: \u001b[36m1/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m8.22s \u001b[35mloss:\u001b[94m 0.376459\u001b[35m acc:\u001b[94m 87.735417\u001b[35m fscore_macro:\u001b[94m 0.876574\u001b[35m val_loss:\u001b[94m 0.089444\u001b[35m val_acc:\u001b[94m 97.358333\u001b[35m val_fscore_macro:\u001b[94m 0.973583\u001b[0m\n",
      "Epoch 1: val_acc improved from -inf to 97.35833, saving file to saves/conv_net_model_bundle/checkpoint_epoch_1.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m2/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m7.81s \u001b[35mloss:\u001b[94m 0.125036\u001b[35m acc:\u001b[94m 96.181250\u001b[35m fscore_macro:\u001b[94m 0.961548\u001b[35m val_loss:\u001b[94m 0.056173\u001b[35m val_acc:\u001b[94m 98.375000\u001b[35m val_fscore_macro:\u001b[94m 0.983650\u001b[0m\n",
      "Epoch 2: val_acc improved from 97.35833 to 98.37500, saving file to saves/conv_net_model_bundle/checkpoint_epoch_2.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m3/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m8.26s \u001b[35mloss:\u001b[94m 0.098956\u001b[35m acc:\u001b[94m 96.979167\u001b[35m fscore_macro:\u001b[94m 0.969604\u001b[35m val_loss:\u001b[94m 0.053156\u001b[35m val_acc:\u001b[94m 98.533333\u001b[35m val_fscore_macro:\u001b[94m 0.985314\u001b[0m\n",
      "Epoch 3: val_acc improved from 98.37500 to 98.53333, saving file to saves/conv_net_model_bundle/checkpoint_epoch_3.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m4/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m8.15s \u001b[35mloss:\u001b[94m 0.086170\u001b[35m acc:\u001b[94m 97.395833\u001b[35m fscore_macro:\u001b[94m 0.973774\u001b[35m val_loss:\u001b[94m 0.045325\u001b[35m val_acc:\u001b[94m 98.675000\u001b[35m val_fscore_macro:\u001b[94m 0.986623\u001b[0m\n",
      "Epoch 4: val_acc improved from 98.53333 to 98.67500, saving file to saves/conv_net_model_bundle/checkpoint_epoch_4.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m5/5 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m8.33s \u001b[35mloss:\u001b[94m 0.075542\u001b[35m acc:\u001b[94m 97.656250\u001b[35m fscore_macro:\u001b[94m 0.976389\u001b[35m val_loss:\u001b[94m 0.040350\u001b[35m val_acc:\u001b[94m 98.816667\u001b[35m val_fscore_macro:\u001b[94m 0.988083\u001b[0m\n",
      "Epoch 5: val_acc improved from 98.67500 to 98.81667, saving file to saves/conv_net_model_bundle/checkpoint_epoch_5.ckpt\n",
      "Restoring data from saves/conv_net_model_bundle/checkpoint_epoch_5.ckpt\n",
      "Found best checkpoint at epoch: 5\n",
      "lr: 0.1, loss: 0.0755423, acc: 97.6562, fscore_macro: 0.976389, val_loss: 0.0403503, val_acc: 98.8167, val_fscore_macro: 0.988083\n",
      "Loading checkpoint saves/conv_net_model_bundle/checkpoint_epoch_5.ckpt\n",
      "Running test\n",
      "\u001b[35mTest steps: \u001b[36m313 \u001b[32m1.14s \u001b[35mtest_loss:\u001b[94m 0.030638\u001b[35m test_acc:\u001b[94m 98.930000\u001b[35m test_fscore_macro:\u001b[94m 0.989233\u001b[0m             \n"
     ]
    }
   ],
   "source": [
    "conv_net = create_convolutional_network()\n",
    "train_model_bundle(conv_net, 'conv_net_model_bundle', epochs=5)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sequential(\n",
      "  (0): Conv2d(1, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (1): ReLU()\n",
      "  (2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (3): Conv2d(16, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
      "  (4): ReLU()\n",
      "  (5): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\n",
      "  (6): Dropout(p=0.25, inplace=False)\n",
      "  (7): Flatten(start_dim=1, end_dim=-1)\n",
      "  (8): Linear(in_features=1568, out_features=128, bias=True)\n",
      "  (9): ReLU()\n",
      "  (10): Dropout(p=0.5, inplace=False)\n",
      "  (11): Linear(in_features=128, out_features=10, bias=True)\n",
      ")\n",
      "Loading weights from saves/conv_net_model_bundle/checkpoint.ckpt and starting at epoch 6.\n",
      "Loading optimizer state from saves/conv_net_model_bundle/checkpoint.optim and starting at epoch 6.\n",
      "\u001b[35mEpoch: \u001b[36m 6/10 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m8.50s \u001b[35mloss:\u001b[94m 0.057106\u001b[35m acc:\u001b[94m 98.158333\u001b[35m fscore_macro:\u001b[94m 0.981433\u001b[35m val_loss:\u001b[94m 0.044382\u001b[35m val_acc:\u001b[94m 98.800000\u001b[35m val_fscore_macro:\u001b[94m 0.987919\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m 7/10 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m8.65s \u001b[35mloss:\u001b[94m 0.047568\u001b[35m acc:\u001b[94m 98.550000\u001b[35m fscore_macro:\u001b[94m 0.985385\u001b[35m val_loss:\u001b[94m 0.042106\u001b[35m val_acc:\u001b[94m 98.758333\u001b[35m val_fscore_macro:\u001b[94m 0.987475\u001b[0m\n",
      "\u001b[35mEpoch: \u001b[36m 8/10 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m8.73s \u001b[35mloss:\u001b[94m 0.043904\u001b[35m acc:\u001b[94m 98.681250\u001b[35m fscore_macro:\u001b[94m 0.986715\u001b[35m val_loss:\u001b[94m 0.041367\u001b[35m val_acc:\u001b[94m 98.825000\u001b[35m val_fscore_macro:\u001b[94m 0.988208\u001b[0m\n",
      "Epoch 8: val_acc improved from 98.81667 to 98.82500, saving file to saves/conv_net_model_bundle/checkpoint_epoch_8.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m 9/10 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m8.71s \u001b[35mloss:\u001b[94m 0.040539\u001b[35m acc:\u001b[94m 98.772917\u001b[35m fscore_macro:\u001b[94m 0.987657\u001b[35m val_loss:\u001b[94m 0.037967\u001b[35m val_acc:\u001b[94m 98.891667\u001b[35m val_fscore_macro:\u001b[94m 0.988824\u001b[0m\n",
      "Epoch 9: val_acc improved from 98.82500 to 98.89167, saving file to saves/conv_net_model_bundle/checkpoint_epoch_9.ckpt\n",
      "\u001b[35mEpoch: \u001b[36m10/10 \u001b[35mTrain steps: \u001b[36m1500 \u001b[35mVal steps: \u001b[36m375 \u001b[32m8.76s \u001b[35mloss:\u001b[94m 0.037301\u001b[35m acc:\u001b[94m 98.808333\u001b[35m fscore_macro:\u001b[94m 0.988002\u001b[35m val_loss:\u001b[94m 0.038459\u001b[35m val_acc:\u001b[94m 98.908333\u001b[35m val_fscore_macro:\u001b[94m 0.989028\u001b[0m\n",
      "Epoch 10: val_acc improved from 98.89167 to 98.90833, saving file to saves/conv_net_model_bundle/checkpoint_epoch_10.ckpt\n",
      "Restoring data from saves/conv_net_model_bundle/checkpoint_epoch_10.ckpt\n",
      "Found best checkpoint at epoch: 10\n",
      "lr: 0.1, loss: 0.0373014, acc: 98.8083, fscore_macro: 0.988002, val_loss: 0.0384585, val_acc: 98.9083, val_fscore_macro: 0.989028\n",
      "Loading checkpoint saves/conv_net_model_bundle/checkpoint_epoch_10.ckpt\n",
      "Running test\n",
      "\u001b[35mTest steps: \u001b[36m313 \u001b[32m0.97s \u001b[35mtest_loss:\u001b[94m 0.026998\u001b[35m test_acc:\u001b[94m 99.140000\u001b[35m test_fscore_macro:\u001b[94m 0.991343\u001b[0m             \n"
     ]
    }
   ],
   "source": [
    "conv_net = create_convolutional_network()\n",
    "train_model_bundle(conv_net, 'conv_net_model_bundle', epochs=10)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Next Step\n",
    "We have plenty more examples for you, see [this](https://github.com/GRAAL-Research/poutyne/blob/master/examples/tips_and_tricks.ipynb) notebook for tips and tricks using Poutyne, this [one](https://github.com/GRAAL-Research/poutyne/blob/master/examples/policy_cifar_example.ipynb) and this [one](https://github.com/GRAAL-Research/poutyne/blob/master/examples/policy_interface.ipynb) for policy examples or this [one](https://github.com/GRAAL-Research/poutyne/blob/master/examples/transfer_learning.ipynb) for a transfer learning example."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "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": "3.8.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}