demos/FashionMNIST_batched.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "db3f8d3e",
   "metadata": {},
   "source": [
    "In this jupyter notebook, we show how EP3.0 can be used with any custom dataloader through fit_batch() module.\n",
    "Here we create a pytorch data loader for Fashion MNIST data set.\n",
    "\n",
    "First we create dataset and dataloader to visualize the dataset:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "311e4f86",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/amirhesam/research/kernel/venv_ker/lib/python3.9/site-packages/torchvision/datasets/mnist.py:498: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at ../torch/csrc/utils/tensor_numpy.cpp:199.)\n",
      "  return torch.from_numpy(parsed.astype(m[2], copy=False)).view(*s)\n"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "from torchvision import datasets, transforms\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "transform = transforms.Compose([transforms.ToTensor()])\n",
    "trainset_tmp = datasets.FashionMNIST('~/', download=True, train=True, transform=transform)\n",
    "trainloader_tmp = torch.utils.data.DataLoader(trainset_tmp, batch_size=10_000, shuffle=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "a6d14b30",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x288 with 12 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_size = 2\n",
    "fig, ax = plt.subplots(3, 4, figsize=(4*plot_size, 2*plot_size))\n",
    "for i in range(12):\n",
    "    image, label = next(iter(trainloader_tmp))\n",
    "    ax[i//4][i%4].imshow(torch.squeeze(image[i,:]),cmap='gray')\n",
    "    ax[i//4][i%4].axis('off')\n",
    "    \n",
    "plt.subplots_adjust(left=0, bottom=0, right=1, top=1, wspace=0.2, hspace=.1)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "db202701",
   "metadata": {},
   "source": [
    "Loading necessary packages from EP3.0:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "30ae4f26",
   "metadata": {},
   "outputs": [],
   "source": [
    "from eigenpro3.utils import accuracy, load_dataset\n",
    "from eigenpro3.datasets import CustomDataset\n",
    "from eigenpro3.models import KernelModel\n",
    "from eigenpro3.kernels import laplacian, ntk_relu\n",
    "from eigenpro3.utils import accuracy,bottomrule, midrule \n",
    "from torch.nn.functional import one_hot\n",
    "import copy"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "c206655f",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Setting up data preconditioner\n",
      "Data: learning rate: 3629.0,batch_size=7258, top eigenvalue:0.12467030435800552, new top eigenvalue:0.00013776824926026165\n",
      "Data preconditioner is ready.\n",
      "Setup time = 11.63 s\n",
      "Done.\n",
      "==================================================\n",
      "Setting up inexact projector\n",
      "Done.\n",
      "-------------------------\n"
     ]
    }
   ],
   "source": [
    "if torch.cuda.is_available():\n",
    "    DEVICES = [torch.device(f'cuda:{i}') for i in range(torch.cuda.device_count())]\n",
    "else:\n",
    "    DEVICES = [torch.device('cpu')]\n",
    "    \n",
    "kernel_fn = lambda x, z: laplacian(x, z, bandwidth=5.0)\n",
    "p = 5_000\n",
    "s = 10_000\n",
    "n_classes = 10\n",
    "\n",
    "\n",
    "centers = trainset_tmp.data[torch.randperm(trainset_tmp.data.shape[0])[:p]].reshape(-1,28*28)/255.0\n",
    "nystrom_samples = trainset_tmp.data[torch.randperm(trainset_tmp.data.shape[0])[:s]].reshape(-1,28*28)/255.0\n",
    "\n",
    "\n",
    "model= KernelModel(n_classes, centers, kernel_fn, nystrom_samples=nystrom_samples,make_dataloader=False,\n",
    "    devices = DEVICES, \n",
    "    multi_gpu=True)\n",
    "\n",
    "\n",
    "##### creating new dataloader using optimal batchsize calculated based on Nystrom\n",
    "transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.1307,), (0.3081,)),transforms.Lambda(lambda x: torch.flatten(x))])\n",
    "trainset = datasets.FashionMNIST('~/', download=True, train=True, transform=transform)\n",
    "trainloader = torch.utils.data.DataLoader(trainset, batch_size=model.batch_size, shuffle=True)\n",
    "testset = datasets.FashionMNIST('~/', download=True, train=False, transform=transform)\n",
    "testloader = torch.utils.data.DataLoader(testset, batch_size=512, shuffle=False)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "a35db218",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Projection SVD_time: 1.01 s, top_q: 741, top_eigval: 0.12, new top_eigval: 4.51e-04\n",
      "Projection setup: Nystrom size=2000, bs_gpu=5000, eta=1109.50, bs=2219, top_eigval=1.24e-01, beta=1.00\n",
      "==================================================\n",
      "epoch    0\t batch    2\n",
      "epoch    0\t batch    4\n",
      "epoch    0\t batch    6\n",
      "epoch    0\t batch    8\n",
      "-------------------------\n",
      "epoch    1        Test accuracy: 68.42%\n",
      "==================================================\n",
      "epoch    1\t batch    2\n",
      "epoch    1\t batch    4\n",
      "epoch    1\t batch    6\n",
      "epoch    1\t batch    8\n",
      "-------------------------\n",
      "epoch    2        Test accuracy: 71.27%\n",
      "==================================================\n",
      "epoch    2\t batch    2\n",
      "epoch    2\t batch    4\n",
      "epoch    2\t batch    6\n",
      "epoch    2\t batch    8\n",
      "-------------------------\n",
      "epoch    3        Test accuracy: 73.07%\n",
      "==================================================\n",
      "epoch    3\t batch    2\n",
      "epoch    3\t batch    4\n",
      "epoch    3\t batch    6\n",
      "epoch    3\t batch    8\n",
      "-------------------------\n",
      "epoch    4        Test accuracy: 73.88%\n",
      "==================================================\n",
      "epoch    4\t batch    2\n",
      "epoch    4\t batch    4\n",
      "epoch    4\t batch    6\n",
      "epoch    4\t batch    8\n",
      "-------------------------\n",
      "epoch    5        Test accuracy: 74.95%\n",
      "==================================================\n",
      "epoch    5\t batch    2\n",
      "epoch    5\t batch    4\n",
      "epoch    5\t batch    6\n",
      "epoch    5\t batch    8\n",
      "-------------------------\n",
      "epoch    6        Test accuracy: 75.56%\n",
      "==================================================\n",
      "epoch    6\t batch    2\n",
      "epoch    6\t batch    4\n",
      "epoch    6\t batch    6\n",
      "epoch    6\t batch    8\n",
      "-------------------------\n",
      "epoch    7        Test accuracy: 76.09%\n",
      "==================================================\n",
      "epoch    7\t batch    2\n",
      "epoch    7\t batch    4\n",
      "epoch    7\t batch    6\n",
      "epoch    7\t batch    8\n",
      "-------------------------\n",
      "epoch    8        Test accuracy: 76.54%\n",
      "==================================================\n",
      "epoch    8\t batch    2\n",
      "epoch    8\t batch    4\n",
      "epoch    8\t batch    6\n",
      "epoch    8\t batch    8\n",
      "-------------------------\n",
      "epoch    9        Test accuracy: 76.72%\n",
      "==================================================\n",
      "epoch    9\t batch    2\n",
      "epoch    9\t batch    4\n",
      "epoch    9\t batch    6\n",
      "epoch    9\t batch    8\n",
      "-------------------------\n",
      "epoch   10        Test accuracy: 77.03%\n",
      "==================================================\n",
      "epoch   10\t batch    2\n",
      "epoch   10\t batch    4\n",
      "epoch   10\t batch    6\n",
      "epoch   10\t batch    8\n",
      "-------------------------\n",
      "epoch   11        Test accuracy: 77.31%\n",
      "==================================================\n",
      "epoch   11\t batch    2\n",
      "epoch   11\t batch    4\n",
      "epoch   11\t batch    6\n",
      "epoch   11\t batch    8\n",
      "-------------------------\n",
      "epoch   12        Test accuracy: 77.39%\n",
      "==================================================\n",
      "epoch   12\t batch    2\n",
      "epoch   12\t batch    4\n",
      "epoch   12\t batch    6\n",
      "epoch   12\t batch    8\n",
      "-------------------------\n",
      "epoch   13        Test accuracy: 77.64%\n",
      "==================================================\n",
      "epoch   13\t batch    2\n",
      "epoch   13\t batch    4\n",
      "epoch   13\t batch    6\n",
      "epoch   13\t batch    8\n",
      "-------------------------\n",
      "epoch   14        Test accuracy: 77.83%\n",
      "==================================================\n",
      "epoch   14\t batch    2\n",
      "epoch   14\t batch    4\n",
      "epoch   14\t batch    6\n",
      "epoch   14\t batch    8\n",
      "-------------------------\n",
      "epoch   15        Test accuracy: 77.94%\n",
      "==================================================\n",
      "epoch   15\t batch    2\n",
      "epoch   15\t batch    4\n",
      "epoch   15\t batch    6\n",
      "epoch   15\t batch    8\n",
      "-------------------------\n",
      "epoch   16        Test accuracy: 78.17%\n",
      "==================================================\n",
      "epoch   16\t batch    2\n",
      "epoch   16\t batch    4\n",
      "epoch   16\t batch    6\n",
      "epoch   16\t batch    8\n",
      "-------------------------\n",
      "epoch   17        Test accuracy: 78.23%\n",
      "==================================================\n",
      "epoch   17\t batch    2\n",
      "epoch   17\t batch    4\n",
      "epoch   17\t batch    6\n",
      "epoch   17\t batch    8\n",
      "-------------------------\n",
      "epoch   18        Test accuracy: 78.40%\n",
      "==================================================\n",
      "epoch   18\t batch    2\n",
      "epoch   18\t batch    4\n",
      "epoch   18\t batch    6\n",
      "epoch   18\t batch    8\n",
      "-------------------------\n",
      "epoch   19        Test accuracy: 78.49%\n",
      "==================================================\n",
      "epoch   19\t batch    2\n",
      "epoch   19\t batch    4\n",
      "epoch   19\t batch    6\n",
      "epoch   19\t batch    8\n",
      "-------------------------\n",
      "epoch   20        Test accuracy: 78.74%\n",
      "==================================================\n",
      "epoch   20\t batch    2\n",
      "epoch   20\t batch    4\n",
      "epoch   20\t batch    6\n",
      "epoch   20\t batch    8\n",
      "-------------------------\n",
      "epoch   21        Test accuracy: 78.82%\n",
      "==================================================\n",
      "epoch   21\t batch    2\n",
      "epoch   21\t batch    4\n",
      "epoch   21\t batch    6\n",
      "epoch   21\t batch    8\n",
      "-------------------------\n",
      "epoch   22        Test accuracy: 78.96%\n",
      "==================================================\n",
      "epoch   22\t batch    2\n",
      "epoch   22\t batch    4\n",
      "epoch   22\t batch    6\n",
      "epoch   22\t batch    8\n",
      "-------------------------\n",
      "epoch   23        Test accuracy: 79.03%\n",
      "==================================================\n",
      "epoch   23\t batch    2\n",
      "epoch   23\t batch    4\n",
      "epoch   23\t batch    6\n",
      "epoch   23\t batch    8\n",
      "-------------------------\n",
      "epoch   24        Test accuracy: 79.12%\n",
      "==================================================\n",
      "epoch   24\t batch    2\n",
      "epoch   24\t batch    4\n",
      "epoch   24\t batch    6\n",
      "epoch   24\t batch    8\n",
      "-------------------------\n",
      "epoch   25        Test accuracy: 79.23%\n",
      "==================================================\n",
      "epoch   25\t batch    2\n",
      "epoch   25\t batch    4\n",
      "epoch   25\t batch    6\n",
      "epoch   25\t batch    8\n",
      "-------------------------\n",
      "epoch   26        Test accuracy: 79.30%\n",
      "==================================================\n",
      "epoch   26\t batch    2\n",
      "epoch   26\t batch    4\n",
      "epoch   26\t batch    6\n",
      "epoch   26\t batch    8\n",
      "-------------------------\n",
      "epoch   27        Test accuracy: 79.49%\n",
      "==================================================\n",
      "epoch   27\t batch    2\n",
      "epoch   27\t batch    4\n",
      "epoch   27\t batch    6\n",
      "epoch   27\t batch    8\n",
      "-------------------------\n",
      "epoch   28        Test accuracy: 79.55%\n",
      "==================================================\n",
      "epoch   28\t batch    2\n",
      "epoch   28\t batch    4\n",
      "epoch   28\t batch    6\n",
      "epoch   28\t batch    8\n",
      "-------------------------\n",
      "epoch   29        Test accuracy: 79.58%\n",
      "==================================================\n",
      "epoch   29\t batch    2\n",
      "epoch   29\t batch    4\n",
      "epoch   29\t batch    6\n",
      "epoch   29\t batch    8\n",
      "-------------------------\n",
      "epoch   30        Test accuracy: 79.67%\n",
      "==================================================\n",
      "epoch   30\t batch    2\n",
      "epoch   30\t batch    4\n",
      "epoch   30\t batch    6\n",
      "epoch   30\t batch    8\n",
      "-------------------------\n",
      "epoch   31        Test accuracy: 79.75%\n",
      "==================================================\n",
      "epoch   31\t batch    2\n",
      "epoch   31\t batch    4\n",
      "epoch   31\t batch    6\n",
      "epoch   31\t batch    8\n",
      "-------------------------\n",
      "epoch   32        Test accuracy: 79.77%\n",
      "==================================================\n",
      "epoch   32\t batch    2\n",
      "epoch   32\t batch    4\n",
      "epoch   32\t batch    6\n",
      "epoch   32\t batch    8\n",
      "-------------------------\n",
      "epoch   33        Test accuracy: 79.86%\n",
      "==================================================\n",
      "epoch   33\t batch    2\n",
      "epoch   33\t batch    4\n",
      "epoch   33\t batch    6\n",
      "epoch   33\t batch    8\n",
      "-------------------------\n",
      "epoch   34        Test accuracy: 79.94%\n",
      "==================================================\n",
      "epoch   34\t batch    2\n",
      "epoch   34\t batch    4\n",
      "epoch   34\t batch    6\n",
      "epoch   34\t batch    8\n",
      "-------------------------\n",
      "epoch   35        Test accuracy: 80.03%\n",
      "==================================================\n",
      "epoch   35\t batch    2\n",
      "epoch   35\t batch    4\n",
      "epoch   35\t batch    6\n",
      "epoch   35\t batch    8\n",
      "-------------------------\n",
      "epoch   36        Test accuracy: 80.03%\n",
      "==================================================\n",
      "epoch   36\t batch    2\n",
      "epoch   36\t batch    4\n",
      "epoch   36\t batch    6\n",
      "epoch   36\t batch    8\n",
      "-------------------------\n",
      "epoch   37        Test accuracy: 80.09%\n",
      "==================================================\n",
      "epoch   37\t batch    2\n",
      "epoch   37\t batch    4\n",
      "epoch   37\t batch    6\n",
      "epoch   37\t batch    8\n",
      "-------------------------\n",
      "epoch   38        Test accuracy: 80.16%\n",
      "==================================================\n",
      "epoch   38\t batch    2\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "epoch   38\t batch    4\n",
      "epoch   38\t batch    6\n",
      "epoch   38\t batch    8\n",
      "-------------------------\n",
      "epoch   39        Test accuracy: 80.23%\n",
      "==================================================\n",
      "epoch   39\t batch    2\n",
      "epoch   39\t batch    4\n",
      "epoch   39\t batch    6\n",
      "epoch   39\t batch    8\n",
      "-------------------------\n",
      "epoch   40        Test accuracy: 80.30%\n",
      "==================================================\n",
      "epoch   40\t batch    2\n",
      "epoch   40\t batch    4\n",
      "epoch   40\t batch    6\n",
      "epoch   40\t batch    8\n",
      "-------------------------\n",
      "epoch   41        Test accuracy: 80.36%\n",
      "==================================================\n",
      "epoch   41\t batch    2\n",
      "epoch   41\t batch    4\n",
      "epoch   41\t batch    6\n",
      "epoch   41\t batch    8\n",
      "-------------------------\n",
      "epoch   42        Test accuracy: 80.38%\n",
      "==================================================\n",
      "epoch   42\t batch    2\n",
      "epoch   42\t batch    4\n",
      "epoch   42\t batch    6\n",
      "epoch   42\t batch    8\n",
      "-------------------------\n",
      "epoch   43        Test accuracy: 80.47%\n",
      "==================================================\n",
      "epoch   43\t batch    2\n",
      "epoch   43\t batch    4\n",
      "epoch   43\t batch    6\n",
      "epoch   43\t batch    8\n",
      "-------------------------\n",
      "epoch   44        Test accuracy: 80.51%\n",
      "==================================================\n",
      "epoch   44\t batch    2\n",
      "epoch   44\t batch    4\n",
      "epoch   44\t batch    6\n",
      "epoch   44\t batch    8\n",
      "-------------------------\n",
      "epoch   45        Test accuracy: 80.56%\n",
      "==================================================\n",
      "epoch   45\t batch    2\n",
      "epoch   45\t batch    4\n",
      "epoch   45\t batch    6\n",
      "epoch   45\t batch    8\n",
      "-------------------------\n",
      "epoch   46        Test accuracy: 80.59%\n",
      "==================================================\n",
      "epoch   46\t batch    2\n",
      "epoch   46\t batch    4\n",
      "epoch   46\t batch    6\n",
      "epoch   46\t batch    8\n",
      "-------------------------\n",
      "epoch   47        Test accuracy: 80.62%\n",
      "==================================================\n",
      "epoch   47\t batch    2\n",
      "epoch   47\t batch    4\n",
      "epoch   47\t batch    6\n",
      "epoch   47\t batch    8\n",
      "-------------------------\n",
      "epoch   48        Test accuracy: 80.68%\n",
      "==================================================\n",
      "epoch   48\t batch    2\n",
      "epoch   48\t batch    4\n",
      "epoch   48\t batch    6\n",
      "epoch   48\t batch    8\n",
      "-------------------------\n",
      "epoch   49        Test accuracy: 80.67%\n",
      "==================================================\n",
      "epoch   49\t batch    2\n",
      "epoch   49\t batch    4\n",
      "epoch   49\t batch    6\n",
      "epoch   49\t batch    8\n",
      "-------------------------\n",
      "epoch   50        Test accuracy: 80.65%\n",
      "==================================================\n"
     ]
    }
   ],
   "source": [
    "iterations = 50\n",
    "epoch = 0\n",
    "accu_all = []\n",
    "for _ in range(iterations):  \n",
    "    batch_num = 0\n",
    "    for (X_batch, y_batch) in trainloader:\n",
    "        X_batch = X_batch \n",
    "        y_batch = one_hot(y_batch)\n",
    "        ###### fitting the batch\n",
    "        model.fit_batch(X_batch, y_batch)\n",
    "        model.sync_gpus()\n",
    "        torch.cuda.empty_cache()\n",
    "        if (batch_num + 1)%2==0:\n",
    "            print(f'epoch {epoch: 4d}\\t batch {batch_num+1 :4d}')\n",
    "        batch_num += 1\n",
    "        del  X_batch, y_batch\n",
    "    \n",
    "    \n",
    "    accu = accuracy(model.weights,model.centers,testloader,model.kernel,model.device_base)\n",
    "    print(midrule)\n",
    "    print(f'epoch {epoch+1:4d}        Test accuracy: {accu*100.:5.2f}%')\n",
    "    accu_all.append(accu.item())\n",
    "    print(bottomrule)\n",
    "    epoch +=1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "e4d0f6e2",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAVAAAAFPCAYAAAAbaJaXAAAAOXRFWHRTb2Z0d2FyZQBNYXRwbG90bGliIHZlcnNpb24zLjQuMywgaHR0cHM6Ly9tYXRwbG90bGliLm9yZy/MnkTPAAAACXBIWXMAAAsTAAALEwEAmpwYAAAuW0lEQVR4nO3deXxU9b3/8deHkBB2AoR9CQjIoogacbfWrVjX3vZa1FZrbb2tS1tvN3tvf9Z6t7Z38ba93ra29aLWulFt0WpBcVeUBAVlUdYgBCEJIZAEsn9+f5wTO45ZJkMmM5N5Px+PeWTOOd8z8zmTyTtn/R5zd0REpOv6JLsAEZF0pQAVEYmTAlREJE4KUBGROClARUTipAAVEYmTAlREJE4KUBGROClAU4yZlZjZOR1MX2dmZ/ZcRSLSHgXoYQoD75CZ1UQ8xiXq/dx9jrs/f7ivE9bdYGYjo8a/aWZuZgUR7crMbGBEmy+Z2fNRr3VOxPBpZvaqme03s0oze8XMToj6jFqiPrcrD3eZusLMnjezuoj3fzfG+Yab2WNmVmtm283silindzZvspnZQjPbENa3xcxOj5jW5doP57OK9z17Wt9kF9BLXOTuzyS7iDhsAy4Hfg5gZkcDA9polwV8HfjXzl7QzIYATwBfBR4GcoDTgXp3HxTRrgT4UpI/txvd/TddnOdOoAEYDcwD/mxma9x9XQzTO5s3aczsXODHwGeBlcDYqCbx1H44n1W879mz3F2Pw3gAJcA5bYy/BdgCVAPrgU9FTf8uUBpOfxc4O+L1vgW8BewHHgJy23o/YBbwPFAFrAMujmrX2et8HyiKGPcfwD8CDhREtLsFqASGheO+BDzfTk2FQFW8n1sb7a4EVoT1vw/sAM7vht/b8wQB3pV5BhL8Qc+IGHcf8KPOpnc2bxdq+DvgSYJwqQB2Aed2w+fxKnBtPMvd3Z9VvO+ZjIc24RNnC8Ga11Dgh8DvzGwsgJkdCdwInODug4FPEARKq8uABcAUYC7whegXN7Ns4HFgGTAKuAm4P3ztWF/nNWCImc0ysyxgIfC7NpalmCBwvhXDcm8Ems3sHjM738zyYpinI0cTrH08BIwHfgr8MrqRmT1hZlXtPJ5o57X/zcwqwl0MZ8ZQywygyd03RoxbA8yJYXpn88bqGOAkYAnB7/1XBP+MP6Qrn0f4uy8E8s1ss5ntNLP/MbP+MS53Ww7ns4r3PXucArR7/DHiy/lHAHd/xN13uXuLuz8EbALmh+2bgX7AbDPLdvcSd98S8Xo/C+etJAjJeW2850nAIIL/yA3u/izBpvPlXXyd+4CrgHOBDQRrxW25FbjJzPI7+iDc/QBwGsFa7K+BcjNbYmajO5qvA0cDd7j7o+7eAtwLTDKz3Kj3vdDdh7XzuLCN1/0uMJUglO8CHjezIzqpZRBwIGrcfmBwDNM7mzdWcwl+50vDz2N9W426+HmMBrKBzxD8058HHEuwhdLZcrXncD6reN+zxylAu8elEV/OSwHM7CozW90arMBRwEgAd98MfAO4DSgzswftwweedkc8P0jwZYo2DtgR/hG12k4QCF15nfuAKwjWTu9tbwHdfS1BQN/SXpuIthvc/QvuPoFguccB/93ZfGZ2ZcRBnafC0UcDiyOajQJq3L2us9frpMbX3b3a3evd/R7gFeCTncxWAwyJGjeEYDdMZ9M7nDd62dv6LMzMCD6PxyNe4yjaCdEuOBT+/Lm7v+/uFcB/8dfPo7PlbsvhfFbxvmePU4AmgJlNJlj7uhEY4e7DgLWAtbZx99+7+2nAZIK1tR938W12ARPNLPJ3OIn21yDb5O7bCQ4mfRJ4tJPmPwC+zIdDurPXfwdYRPCH3lnb+919UPg438yGAROB8ohmnwGeip43DJ2adh4fad/W2xPx+2nHRqCvmU2PGHcMwf7nzqZ3OG/0skcPh+0LCA78Rp4xcCywOrrQrnwe7r4P2Bl+BpGfR6zL3ZbD+azifc8epwBNjIEEX8ByADO7hogAMbMjzewsM+sH1BGsAbS09UIdeJ1grfI7ZpYd7sO7CHgwjnqvBc5y99qOGoVrzg8BX2uvjZnNNLNvmtmEcHgiwW6F1+Ko62iC3R1XmFlfM7sAuJ5gzT26tvMjAif6cX5kWzMbZmafMLPc8HWvBM4A/hLRZpGZLYp6j1qCfzK3m9lAMzsVuIRgLb7D6Z3NG6O5wNtRWx3HEuwbjPvzCP0fwS6aUeF+65sJtjg6Xe7u/qximZ4qFKAJ4O7rgf8kOHq8hyAIXolo0o/gyGwFwWb2KOB7XXyPBoLAPD98nf8FrgrX+Lpa7xZ3L46x+e0E/yDaUw2cCLxuZrUEwbkW+GZX6yL43O4HTgb2ERyMuzT8fA9HNvDPBP/gKggOwF0adcBiIh/+nbW6HugPlAEPAF/1D59W09H0zubtzFwi1jYtOId3DMHne7j+CSgiWPPbALwJ/EvE9I5qT8RnFcv0pDN33dJDUpOZ/QLY6O539PD75hCs1c1198aefO90k+mfldZAJZUdTbA21KPCsxpmZWIgdFWmf1YKUEllRwFd3iUh0lO0CS8iEietgYqIxEkBKiISp17TG9PIkSO9oKAg2WWISC+zatWqCndv8xLmXhOgBQUFFBfHeiqjiEhszGx7e9O0CS8iEicFqIhInBSgIiJxUoCKiMRJASoiEicFqIhInBSgIiJxUoCKiMRJASoiEqdecyWSiGSepuYW1r9/gHd2V5OT1Yfc7Cz652TRPzuLATlZDO2fTd7AHAbmZBHck697KUBFJC00tzjl1fVsKqumqGQfxSWVrN5RxcGG5k7nzcnqQ97AbPIG5HDliZP4/MkF3VKTAlREUkJ1XSO7qurYVXWI0qpD7PrgUUdp1SH2HKijqSXov7iPwayxQ7iscCLHT87jqPFDcXfqGls41NhMXWMztfVNVB1qZF9tA/sOBj8rDzYwIKf7Yk8BKiI9yt15r/Iga0sPsHbXftaW7mfdrgNU1jZ8qF3fPsboIbmMz+vP/CnDGTcsl3HD+jN5+EDmTRrGoH7Jj6/kVyAivZK7s6W8lk17qtlSXsPW8lq2lNewpbyWmvomALKzjBmjB3POrFFMzR/E+GH9GTesP+OH9Sd/cD+y+nT/fsvupAAVkW61tbyGP63exeNrdrG1ovaD8eOG5nLEqEF8+rjxzBo7hKPGD2X66EH065uVxGoPjwJURLrM3ampb2JfbSOVBxvYV9vAprJqHl/zPm+X7scMTpoygmtPn8IxE4YxNX9gt+57TBW9b4lEpNu9v/8QRSX7WFVSSVHJPjaX1dDQ3PKRdnMnDOX7F8ziwrnjGDM0NwmV9iwFqIh8RHl1PS9vLuelTRW8vrWS0qpDAAzIyeK4SXlcc2oBIwblkDcgh+EDc8gbmMOYIcFBnkyiABUR6hqbKS7Zx0ubynlxUwUb3j8AQN6AbE6aOoJrT5vCCQXDmTV2MH2zdAFjKwWoSAZyd97dU81LGyt4cVM5K7dVUt/UQnaWcfzkPL79iSM5Y3o+c8YNoU+KHwlPpoQGqJktAH4KZAG/cfcfRU2fBNwDDAvb3OLuT4bTvgdcCzQDX3P3pYmsVaS3i9wsf3lTBWXV9QBMGzWIK06cxBnT85k/ZTgDU+D8ynSRsE/KzLKAO4FzgZ1AkZktcff1Ec2+Dzzs7r8ws9nAk0BB+HwhMAcYBzxjZjPcvfNrtkQEgKqDDazavo/Xt1XyUtRm+anTRnLG9HxOmz4y4/ZbdqdE/quZD2x2960AZvYgcAkQGaAODAmfDwV2hc8vAR5093pgm5ltDl9vRQLrFUkbOyoPsnFP9UfG7zvYyKrtwXXim8pqgOBk9cLJw/nOgiM5fZo2y7tTIgN0PLAjYngncGJUm9uAZWZ2EzAQOCdi3tei5h2fmDJF0kN5dT1/fmsXf1qzizffq2q33eDcvhROzuPSY8dTODmPYyYOIzc7fU9WT2XJ3tlxObDI3f/TzE4G7jOzo2Kd2cyuA64DmDRpUoJKFEmevTX1LN9QxuNv7eKVzRW0OMwcM5jvLpjJSVOHf+RSxwE5WUwdOUhrmD0kkQFaCkyMGJ4Qjot0LbAAwN1XmFkuMDLGeXH3u4C7AAoLC73bKhdJou17a1m2bg9Pr99D8fZKWhwmDu/P9WdO4+J545gxenCyS5RQIgO0CJhuZlMIwm8hcEVUm/eAs4FFZjYLyAXKgSXA783svwgOIk0HViawVpGk2lF5kCVrguvH39kd7NucOWYwN541nfNmj2bOuCEJ6RBYDk/CAtTdm8zsRmApwSlKd7v7OjO7HSh29yXAN4Ffm9nNBAeUvuDuDqwzs4cJDjg1ATfoCLz0Ju5B58BPrd3NkjW7WLV9HwDHT87j+xfM4hNzxjBx+IAkVymdsSCv0l9hYaEXFxcnuwyRDyk7UMfTG/bw2tZK9tbUU1nbQGVtA/sONtDYHPztzRwzmIvnjeOiueMUminIzFa5e2Fb05J9EEmk19lcVsOy9btZtm4Pq3dUAUFXbuOG9Wfi8AEcM2EYeQNzGDkoh9On53PkGO3TTFcKUJFu0LoPc8nqXbwbnp95zIShfOu8GZw3ZwzTRw3SPsxeSAEqEoea+iZ2VR3i1c0VLFmzizfC8zILJ+fxw4vncN6c0Ywdqit8ejsFqEgHyqvrWbW9kuKSfWyrqP3gZmcH6po+aDNr7BC+u2AmFx0zlgl52oeZSRSgIhHcnSfeep8XN5ZTvD0ITYB+ffswNX8QE/IGhDc4C+7dM3vsYKaN0j7MTKUAFQntq23g24vX8MyGMvIGZFNYMJzL50+ksGA4R40bSk5f9YMpH6YAFQGKSir52gNvsremgdsums3VpxTooI90SgEqGa25xfnF85u545lNTMzrz6PXn8JR44cmuyxJEwpQyUhVBxt4dcte7luxnRVb93LxMeP4l08dxeDc7GSXJmlEASoZobnFWbX9r/f8eXtnFS0OQ3L78uNPH81lhRO1yS5dpgCVXm3/oUYeKd7BvSu2817lQfoYHDspj5vOms4ZM0ZyzIRhukmaxE0BKr3S5rJqFr1awqNvlHKwoZkTCvL45nkzOPPIUQztr8106R4KUOk19tbU8+Ta3SxZXUpRyT5y+vbh4mPG8YVTCnRgSBJCASpprbqukafX72HJml28tKmC5hZn+qhBfGfBkXy2cCIjBvVLdonSiylAJa00tzhv7azipU0VvLSpnDfeq6K5xRk/rD/XnTGVi48Zx8wxg3VASHqEAlTSwva9tfxs+Wae2bCH/YcaMYOjxw/l786YylkzR3HcpDzdB0h6nAJUUlplbQM/W76J+1/fTt8+fbhw7ljOmJHPqdNGMnxgTrLLkwynAJWUdKihmbtf2cYvn9/CwcZmLiucyM3nTGfUkNxklybyAQWopIyGphZe27qXZet385e1e6ioqefc2aP57oIj1eORpCQFqCRVS4vz1Nrd/GXdbp5/p4zq+ib6Z2fxsRn5fPG0KcyfMjzZJYq0SwEqSdPU3MJ3Fr/Fo2+WMmJgDp88eiznzRnNqdNGkpudlezyRDqlAJWkaGxu4eaHVvPEW+/z9+fO4IaPTyNLR9ElzShApcc1NLVw0wNvsHTdHv7hkzO57owjkl2SSFwUoNKj6hqbuf7+N3j2nTJ+cNFsrjl1SrJLEombAlR6zIG6Rm64/w1e2lTBP196FJ87aXKySxI5LApQSai6xmaee6eMP63exbPvltHY3MJPPj2Xy06YmOzSRA6bAlQSYm3pfha9WsLStbuprm9i5KB+XDF/Ep85foJ6RpJeQwEq3aq+qZmfLd/EL1/YyoDsLBYcNYZL5o3npKnD1XGx9DoKUOk2q3dU8e1H1rCprIbPHD+B/3fBbIYOUOfF0nspQOWw1TU2c8czG/n1i1sZNTiX/7vmBD5+5KhklyWScApQ6bK6xmbeLt1PUUklxSX7KC6p5EBdEwtPmMg/XDCLIbqzpWQIBajEbG9NPd9Z/BYvbaqgobkFgCPyB/LJo8dy8THjOGXayCRXKNKzEhqgZrYA+CmQBfzG3X8UNf0O4OPh4ABglLsPC6f9BLgA6AM8DXzd3T2R9Ur7dlQe5Kq7V7Kr6hBXnTyZE6eO4PjJeeqTUzJawgLUzLKAO4FzgZ1AkZktcff1rW3c/eaI9jcBx4bPTwFOBeaGk18GPgY8n6h6pX1rS/dzzaIiGppauP9LJ1JYoB6SRCBYu0uU+cBmd9/q7g3Ag8AlHbS/HHggfO5ALpAD9AOygT0JrFXa8fKmChbe9RrZfYzFXzlZ4SkSIZEBOh7YETG8Mxz3EWY2GZgCPAvg7iuA54D3w8dSd9/QxnzXmVmxmRWXl5d3c/nyp9WlXLNoJRPy+vPo9acyfbQ6NRaJlCoHkRYCi929GcDMpgGzgAnh9KfN7HR3fylyJne/C7gLoLCwUPtHu0FjcwtPrd3NPa+WsGr7Pk6cMpy7ripkaH8dWReJlsgALQUiL3ieEI5ry0LghojhTwGvuXsNgJk9BZwMvNTGvNINKmrqeeD19/jd69vZc6CeySMGcOuFs7nixEnq3FikHYkM0CJguplNIQjOhcAV0Y3MbCaQB6yIGP0e8GUz+zfACA4g/XcCa81YLS3OXS9t5b+WbaShuYUzZuTzo78p4GMz8nWbYJFOJCxA3b3JzG4ElhKcxnS3u68zs9uBYndfEjZdCDwYdYrSYuAs4G2CA0p/cffHE1VrptpbU8/fP7yGFzaWs2DOGL694EiOyB+U7LJE0ob1llMrCwsLvbi4ONllpI1Xt1TwjQdXU3WokVsvnM2VJ07CTGucItHMbJW7F7Y1LVUOIkkPaW5xfrp8Ez9/dhNTRg5k0TXzmT1uSLLLEklLCtAMUt/UzE2/f5Nl6/fw6eMmcPslcxjYT18BkXjprydD1DU289XfreK5d8u59cLZfPE03YtI5HApQDPAoYZmrruvmJc2VfCvnzqaK06clOySRHoFBWgvV1vfxLX3FPH6tkp+8pm5XFaoexGJdBcFaC9WXdfINf9XxBvv7eOOy+Zx6bFtXkkrInFSgPYytfVNvLixnGXr97B8wx4ONjTz88uP44K5Y5NdmkivowDtJZ5Zv4ffr3yPlzdX0NDUQt6AbM6dPYaF8ydygnpQEkkIBWgv8KfVpXzjodWMG9qfz504mfPmjKZwcp7ugimSYArQNPfCxnK++fAaTigYzr1fnK+OP0R6kFZR0tib7+3jK/etYvrowfzm6kKFp0gPU4CmqU17qrlmURGjhvTjni+eoDthiiSBAjQNlVYd4qq7V5Kd1Yf7vngiowbnJrskkYykAE0z+w82ctVvX6emvol7vzifSSMGJLskkYylAE0jTc0t3PD7N3iv8iC/vqqQWWPVi5JIMukofBr55z9v4OXNFfzk03M5aeqIZJcjkvG0Bpom7n99O4teLeFLp03hshN0PbtIKlCApoEVW/bygz+t42Mz8vneJ2cluxwRCSlAU9x7ew/y1ftXMXnEAH5+xbFk6UZvIilDAZrC9hyo49p7inCH316tcz1FUo0OIqWop95+n+899jZ1jc3cffUJFIwcmOySRCSKAjTFVNc1ctuS9fzhjZ3MnTCUOz47T7caFklRCtAUUlRSyc0PrWZX1SFuOmsaXzt7OtnqUUkkZSlAU8Sydbv5yu9WMT6vP4985WSOn6w+PEVSnQI0BdQ1NvPDx9czY/RgFn/1FAbpVsMiaUHbhyngty9vo7TqELdeNFvhKZJGFKBJVnagjv99bjPnzR7NKUeMTHY5ItIFCtAk+49l79LQ3MI/6AojkbSjAE2itaX7eWTVTr546hSd5ymShhSgSeLu3P74eoYPyOGGs6YluxwRiYMCNEmeWrublSWVfPO8I3WJpkia6jRAzewiM4sraM1sgZm9a2abzeyWNqbfYWarw8dGM6uKmDbJzJaZ2QYzW29mBfHUkIrqGpv51yc3MHPMYD6rrulE0lYswfhZYJOZ/cTMZsb6wmaWBdwJnA/MBi43s9mRbdz9Znef5+7zgJ8Dj0ZMvhf4d3efBcwHymJ971R3z6sl7Nx3iFsvnK3elUTSWKcB6u6fA44FtgCLzGyFmV1nZoM7mXU+sNndt7p7A/AgcEkH7S8HHgAIg7avuz8d1lDj7gc7X5zUd7ChiV+9uJWPzcjnlGk6bUkkncW0ae7uB4DFBCE4FvgU8IaZ3dTBbOOBHRHDO8NxH2Fmk4EpwLPhqBlAlZk9amZvmtm/h2u0ae+BlTuorG3ga2frwJFIuotlH+jFZvYY8DyQDcx39/OBY4BvdlMdC4HF7t4cDvcFTge+BZwATAW+0EZt15lZsZkVl5eXd1MpiVPX2MxdL27h5KkjdK27SC8Qyxrop4E73P1od/93dy8DCDepr+1gvlIg8gjJhHBcWxYSbr6HdgKrw83/JuCPwHHRM7n7Xe5e6O6F+fn5MSxKcj2yaid7DtRzk05bEukVYgnQ24CVrQNm1r/1iLi7L+9gviJguplNMbMcgpBcEt0oPDCVB6yImneYmbWm4lnA+hhqTVmNzS388vktHDdpGCcfoTtqivQGsQToI0BLxHBzOK5D4ZrjjcBSYAPwsLuvM7PbzeziiKYLgQfd3SPmbSbYfF9uZm8DBvw6hlpT1mNvllJadYibzpqOmY68i/QGsXT90zc8ig6AuzeEa5SdcvcngSejxt0aNXxbO/M+DcyN5X1SXXOL87/PbWbOuCGceWTq72oQkdjEsgZaHrnGaGaXABWJK6n3eeKtXZTsPchNZ03T2qdILxLLGuhXgPvN7H8INqV3AFcltKpepKXFufO5zUwfNYjzZo9Jdjki0o06DVB33wKcZGaDwuGahFfViyxbv4eNe2r46cJ59NFVRyK9Skzdn5vZBcAcILd1E9Tdb09gXb3GL17YQsGIAVxw9NhklyIi3SyWE+l/SXA9/E0Em/B/C0xOcF29wuodVazZUcU1p06hr+6uKdLrxPJXfYq7XwXsc/cfAicTXGopnbh3RQkDc7L4m+PavIJVRNJcLAFaF/48aGbjgEaC6+GlA3tr6nlizfv8zXETGKz+PkV6pVj2gT5uZsOAfwfeAJw0P6m9JzxUvIOG5hauOll7O0R6qw4DNOxIebm7VwF/MLMngFx3398TxaWrpuYW7n/tPU45YgTTR3fW65+IpKsON+HdvYWgU+TW4XqFZ+eWv1NGadUhrjq5INmliEgCxbIPdLmZfdp0CU3M7l1RwrihuZwza1SySxGRBIolQP+OoPOQejM7YGbVZnYgwXWlrc1l1byyeS9XnjRZpy6J9HKxXImknXhdcN+K7eRk9dHN4kQyQKcBamZntDXe3V/s/nLSW3VdI4tX7eTCuWMZOahfsssRkQSL5TSmb0c8zyW4Wdwqgk6OJcJjb5ZS29DMVacUJLsUEekBsWzCXxQ5bGYTgf9OVEHp7P7X3uOYCUOZN3FYsksRkR4Qz1GOncCs7i4k3ZUdqOPdPdVcOHdcsksRkR4Syz7QnxNcfQRB4M4juCJJIqwsqQRg/hTdbVMkU8SyD7Q44nkT8IC7v5KgetJW0bZKBuRkMWfckGSXIiI9JJYAXQzUtd6z3cyyzGxAeFtjCb2+rZLjJ+fp3E+RDBLTlUhA/4jh/sAziSknPe0/1Mi7e6o5oUCb7yKZJJYAzY28jUf4fEDiSko/q7ZX4q79nyKZJpYArTWz41oHzOx44FDiSko/r2+rJDvLdPqSSIaJZR/oN4BHzGwXwS09xhDc4kNCRdsqOWbCMHKzs5Jdioj0oFhOpC8ys5nAkeGod929MbFlpY9DDc28tXM/Xz5jarJLEZEeFstN5W4ABrr7WndfCwwys+sTX1p6eHPHPppanPk6gCSScWLZB/rlsEd6ANx9H/DlhFWUZlZuq8QMji/IS3YpItLDYgnQrMjOlM0sC8hJXEnppaikklljhjBEN44TyTixBOhfgIfM7GwzOxt4IByX8RqbW3hje5VOXxLJULEchf8uQa/0Xw2HnwZ+k7CK0sja0v0camxWgIpkqFiOwrcAvwgfEmHltqADEV2BJJKZYjkKP93MFpvZejPb2vqI5cXNbIGZvWtmm83sljam32Fmq8PHRjOripo+xMx2mtn/xLxEPaiopJKpIweSP1i9z4tkolg24f8P+AFwB/Bx4BpiC94sglsin0vQh2iRmS1x9/Wtbdz95oj2NwHHRr3MPwEpeeuQlhanqGQf5x81JtmliEiSxHIQqb+7LwfM3be7+23ABTHMNx/Y7O5b3b0BeBC4pIP2lxMcoAI+uGR0NLAshvfqcRvLqtl/qFGb7yIZLJYArTezPsAmM7vRzD4FDIphvvHAjojhneG4jzCzycAU4NlwuA/wn8C3YnifpGjd/6kDSCKZK5YA/TpB70tfA44HPgdc3c11LAQWt/Y5ClwPPOnuOzuaycyuM7NiMysuLy/v5pI6tnJbJWOH5jIhr3/njUWkV4rpWvjwaQ3B/s9YlQKRN0efEI5ry0Lghojhk4HTw0tGBwE5Zlbj7h86EOXudwF3ARQWFjo9xN1Zua2Sk6aOIOIaAxHJMLEcRIpXETDdzKYQBOdC4IroRmFHJXnAitZx7n5lxPQvAIXR4ZlMW8prKKuu58Sp2nwXyWQJu/+EuzcBNwJLgQ3Aw+6+zsxuN7OLI5ouBB509x5bgzxcS9ftAeDsmaOTXImIJFMsd+U8Nfomcm2Na4u7Pwk8GTXu1qjh2zp5jUXAos7eqyctXbebeROHMWZobrJLEZEkimUN9OcxjssIpVWHeGvnfj4xR+d/imS6dtdAzexk4BQg38z+PmLSECBju15ftm43AJ+Yo813kUzX0SZ8DsER8L7A4IjxB4DPJLKoVLZ03W6mjxrE1PxYToUVkd6s3QB19xeAF8xskbtvhw9OcB/k7gd6qsBUUlnbwMptlVx/5rRklyIiKSCWfaD/FnbqMRBYC6w3s28nuK6U9MyGPbQ4LND17yJCbAE6O1zjvBR4iuCSy88nsqhUtWzdbsYP68+ccUOSXYqIpIBYAjTbzLIJAnRJeEfOtDlns7vU1Dfx4qYKzpszWlcfiQgQW4D+CigBBgIvhh1/ZNw+0BfeLaehqUWnL4nIB2K5Fv5nwM8iRm03s48nrqTUtHTdboYPzFH3dSLygVg6Rh5tZr81s6fC4dl0f29MKa2hqYXn3injnFmjyOqjzXcRCcSyCb+I4Hr2ceHwRuAbCaonJb26pYLq+iYdfReRD2k3QM2sdfN+pLs/DLTAB52ENLc3X2+0dN1uBuZkccoRI5NdioikkI7WQFeGP2vNbAThkXczOwnYn+jCUkVzi/P0+j2cOXMUudkZewWriLSho4NIrTv7/h5YAhxhZq8A+WTQpZxv7ayioqZBR99F5CM6CtDITkQeI+iWzoB64BzgrQTXlhI27qkGYN6EYcktRERSTkcBmkXQmUj0YecBiSsn9ZTsPUh2ljFumPr+FJEP6yhA33f323uskhRVUlHLxLwB9M1KWOf9IpKmOkoFnfBIsAZaMHJgsssQkRTUUYCe3WNVpCh3Z/veWiaPyKi9FiISo3YD1N0re7KQVFRWXc/BhmamaA1URNqgHXsdKKmoBWDyCAWoiHyUArQDJXuDAJ2iABWRNihAO6BTmESkIwrQDugUJhHpiJKhAzqFSUQ6ogBth05hEpHOKEDbUa5TmESkEwrQdmzTKUwi0gkFaDu27z0I6BQmEWmfArQd2/bW0rePTmESkfYpQNuxfW8tk4brFCYRaV9C08HMFpjZu2a22cxuaWP6HWa2OnxsNLOqcPw8M1thZuvM7C0z+2wi62zLtoqDOgIvIh3q9L7w8TKzLOBO4FxgJ1BkZkvcfX1rG3e/OaL9TcCx4eBB4Cp332Rm44BVZrbU3asSVW+k1lOYTpqqe8CLSPsSuQY6H9js7lvdvQF4ELikg/aXAw8AuPtGd98UPt8FlBHci6lHtJ7CVKADSCLSgUQG6HhgR8TwznDcR5jZZGAK8Gwb0+YDOcCWBNTYptZTmHQVkoh0JFWOkCwEFrv7h+43b2ZjgfuAa9y9JXomM7vOzIrNrLi8vLzbimk9halA+0BFpAOJDNBSYGLE8IRwXFsWEm6+tzKzIcCfgX9099famsnd73L3QncvzM/vvi381lOYxg/r322vKSK9TyIDtAiYbmZTzCyHICSXRDcys5lAHrAiYlwOwa2U73X3xQmssU3b99YyUacwiUgnEpYQ7t4E3AgsBTYAD7v7OjO73cwujmi6EHjQ3T1i3GXAGcAXIk5zmpeoWqNtqziozXcR6VTCTmMCcPcngSejxt0aNXxbG/P9DvhdImtrT+spTCdO0SlMItIxbaNGUS9MIhIrBWiUkvAIvK5CEpHOKECjtN6JU2ugItIZBWiUEp3CJCIxUoBGKdEpTCISI6VElBL1wiQiMVKARnB3SvbWqhMREYmJAjRCeY1OYRKR2ClAI5RU6BQmEYmdAjRCyV6dwiQisVOARni/qg6AsUN1CpOIdE4BGqGipp6h/bPJ6auPRUQ6p6SIUF5dT/7gfskuQ0TShAI0QkVNPSMH5SS7DBFJEwrQCEGAag1URGKjAI1QUdOgABWRmClAQ3WNzdTUN2kfqIjETAEaKq+uB9A+UBGJmQI0VFHTGqBaAxWR2ChAQxU1DYACVERipwANfbAGqn2gIhIjBWioItwHOmKg9oGKSGwUoKGKmnoG5/YlNzsr2aWISJpQgIYqahrI1/5PEekCBWioXFchiUgXKUBDFTX1jBys/Z8iEjsFaKi8ul6b8CLSJQpQgss4q+uatAkvIl2iAAX21oYn0escUBHpAgUofz0HVGugItIVClAir4PXQSQRiZ0CFHUkIiLxSWiAmtkCM3vXzDab2S1tTL/DzFaHj41mVhUx7Woz2xQ+rk5kna0diagvUBHpir6JemEzywLuBM4FdgJFZrbE3de3tnH3myPa3wQcGz4fDvwAKAQcWBXOuy8RtZZX1zOony7jFJGuSeQa6Hxgs7tvdfcG4EHgkg7aXw48ED7/BPC0u1eGofk0sCBRhepmciISj0QG6HhgR8TwznDcR5jZZGAK8GxX5+0OupmciMQjVQ4iLQQWu3tzV2Yys+vMrNjMisvLy+N+c91MTkTikcgALQUmRgxPCMe1ZSF/3XyPeV53v8vdC929MD8/P+5CdR28iMQjkQFaBEw3sylmlkMQkkuiG5nZTCAPWBExeilwnpnlmVkecF44rts1NLVQdbBRa6Ai0mUJOwrv7k1mdiNB8GUBd7v7OjO7HSh299YwXQg86O4eMW+lmf0TQQgD3O7ulYmoc29tcA6oTmESka5KWIACuPuTwJNR426NGr6tnXnvBu5OWHGhimrdTE5E4pMqB5GSRlchiUi8Mj5Ay8MAVV+gItJVGR+gf72dsY7Ci0jXKECrGxiQk8WAnITuDhaRXkgBqquQRCROClBdBy8icVKAag1UROKkAK1p0L2QRCQuGR2gTc0t7DuojkREJD4ZHaCVtQ24Q772gYpIHDI6QMt1FZKIHIaMDtDWeyFpH6iIxCOjA7S8Wpdxikj8MjpA/3oZpwJURLouswO0up7c7D4MzNHdOEWk6zI7QMOT6M0s2aWISBrK8ADVOaAiEr8MD1Bdxiki8cv4AM1XP6AiEqeMDdDmFqeyVpvwIhK/jA3QytoGWlxXIYlI/DI2QHUzORE5XApQdSQiInFSgOoqJBGJU+YGaHXYkYg24UUkTpkboDX15GT1YUiu7sYpIvHJ2AAtr64nf7Au4xSR+GXs6td3FsyksrYh2WWISBrL2AAdMzSXMUNzk12GiKSxjN2EFxE5XApQEZE4JTRAzWyBmb1rZpvN7JZ22lxmZuvNbJ2Z/T5i/E/CcRvM7Gemoz0ikmIStg/UzLKAO4FzgZ1AkZktcff1EW2mA98DTnX3fWY2Khx/CnAqMDds+jLwMeD5RNUrItJViVwDnQ9sdvet7t4APAhcEtXmy8Cd7r4PwN3LwvEO5AI5QD8gG9iTwFpFRLoskQE6HtgRMbwzHBdpBjDDzF4xs9fMbAGAu68AngPeDx9L3X1DAmsVEemyZJ/G1BeYDpwJTABeNLOjgZHArHAcwNNmdrq7vxQ5s5ldB1wHMGnSpJ6qWUQESOwaaCkwMWJ4Qjgu0k5gibs3uvs2YCNBoH4KeM3da9y9BngKODn6Ddz9LncvdPfC/Pz8hCyEiEh7EhmgRcB0M5tiZjnAQmBJVJs/Eqx9YmYjCTbptwLvAR8zs75mlk1wAEmb8CKSUhIWoO7eBNwILCUIv4fdfZ2Z3W5mF4fNlgJ7zWw9wT7Pb7v7XmAxsAV4G1gDrHH3xxNVq4hIPMzdk11DtzCzcmB7F2cbCVQkoJxk0fKkrt60LJBZyzPZ3dvcR9hrAjQeZlbs7oXJrqO7aHlSV29aFtDytNKlnCIicVKAiojEKdMD9K5kF9DNtDypqzctC2h5gAzfByoicjgyfQ1URCRuGRugsXS1l8rM7G4zKzOztRHjhpvZ02a2KfyZl8waY2VmE83suYhuDb8ejk/X5ck1s5VmtiZcnh+G46eY2evhd+6h8AKTtGBmWWb2ppk9EQ6n87KUmNnbZrbazIrDcXF91zIyQCO62jsfmA1cbmazk1tVly0CFkSNuwVY7u7TgeXhcDpoAr7p7rOBk4Abwt9Hui5PPXCWux8DzAMWmNlJwI+BO9x9GrAPuDZ5JXbZ1/nw1YDpvCwAH3f3eRGnLsX1XcvIACW2rvZSmru/CFRGjb4EuCd8fg9waU/WFC93f9/d3wifVxP8oY4nfZfHwz4cIOiKMZugi8azCK6ygzRaHjObAFwA/CYcNtJ0WToQ13ctUwM0lq720tFod38/fL4bGJ3MYuJhZgXAscDrpPHyhJu8q4Ey4GmCS5OrwkucIb2+c/8NfAdoCYdHkL7LAsE/s2Vmtirs0Q3i/K4luzs7SRB3dzNLq1MszGwQ8AfgG+5+IPIuLum2PO7eDMwzs2HAY8DM5FYUHzO7EChz91VmdmaSy+kup7l7aXgHjKfN7J3IiV35rmXqGmgsXe2loz1mNhYg/FnWSfuUEfa69Qfgfnd/NBydtsvTyt2rCDrKORkYZmatKy3p8p07FbjYzEoIdnWdBfyU9FwWANy9NPxZRvDPbT5xftcyNUBj6WovHS0Brg6fXw38KYm1xCzcp/ZbYIO7/1fEpHRdnvxwzRMz609wX7ANBEH6mbBZWiyPu3/P3Se4ewHB38mz7n4labgsAGY20MwGtz4HzgPWEu93zd0z8gF8kqAD5y3APya7njjqf4DgdieNBPugriXYN7Uc2AQ8AwxPdp0xLstpBPul3gJWh49PpvHyzAXeDJdnLXBrOH4qsBLYDDwC9Et2rV1crjOBJ9J5WcK614SPda1/+/F+13QlkohInDJ1E15E5LApQEVE4qQAFRGJkwJURCROClARkTgpQEXaYGZntvY8JNIeBaiISJwUoJLWzOxzYd+bq83sV2EnHjVmdkfYF+dyM8sP284zs9fM7C0ze6y1z0czm2Zmz4T9d75hZkeELz/IzBab2Ttmdr9FXpwvggJU0piZzQI+C5zq7vOAZuBKYCBQ7O5zgBeAH4Sz3At8193nAm9HjL8fuNOD/jtPIbjCC4Jeob5B0GfsVILrwkU+oN6YJJ2dDRwPFIUrh/0JOoFoAR4K2/wOeNTMhgLD3P2FcPw9wCPhddHj3f0xAHevAwhfb6W77wyHVwMFwMsJXypJGwpQSWcG3OPu3/vQSLP/F9Uu3uuV6yOeN6O/F4miTXhJZ8uBz4T9Orbe12Yywfe6taegK4CX3X0/sM/MTg/Hfx54wYMe8Hea2aXha/QzswE9uRCSvvQfVdKWu683s+8T9C7eh6BnqhuAWmB+OK2MYD8pBN2U/TIMyK3ANeH4zwO/MrPbw9f42x5cDElj6o1Jeh0zq3H3QcmuQ3o/bcKLiMRJa6AiInHSGqiISJwUoCIicVKAiojESQEqIhInBaiISJwUoCIicfr/+6L9+NNFnGEAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<Figure size 360x360 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_size = 5\n",
    "fig, ax = plt.subplots(1, 1, figsize=(plot_size, plot_size))\n",
    "ax.plot(range(len(accu_all)),accu_all)\n",
    "ax.set_xlabel('epoch')\n",
    "ax.set_ylabel('Test accuracy')\n",
    "ax.set_title(r'FashionMNIST-$p=5,000$--$n=60,000$')\n",
    "plt.show()\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "27a0056e",
   "metadata": {},
   "source": [
    "At the end, please note that the PyTorch DataLoader can be slow, and we recommend using simple batching techniques with manual methods instead. We provided the DataLoader based on PyTorch because it is widely used, but for large-scale experiments, we suggest avoiding it."
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.9.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}