examples.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "import tensorflow as tf\n",
    "import os\n",
    "\n",
    "os.environ['CUDA_VISIBLE_DEVICES'] = '-1'\n",
    "np.set_printoptions(linewidth=120, precision=4)\n",
    "\n",
    "from keras.layers import Input, Conv2D, Activation, Reshape, Lambda\n",
    "from keras.layers import MaxPool2D, AvgPool2D, MaxPool3D, AvgPool3D\n",
    "from keras.layers import BatchNormalization, Dropout\n",
    "from keras.models import Model\n",
    "from keras import backend as K"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### SparseConv2D and PartialConv2D\n",
    "\n",
    "see [sparseconvolution.ipynb](sparseconvolution.ipynb)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### GroupConv2D"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(None, 8, 8, 2)\n",
      "(None, 4, 4, 12)\n",
      "input_1                        [(None, 8, 8, 2)]        0        []\n",
      "group_conv2d                   (None, 8, 8, 8, 12)      216      [(3, 3, 2, 12)]\n",
      "batch_normalization            (None, 8, 8, 8, 12)      48       [(12,), (12,), (12,), (12,)]\n",
      "group_conv2d_1                 (None, 8, 8, 8, 12)      10368    [(3, 3, 96, 12)]\n",
      "average_pooling3d              (None, 4, 4, 8, 12)      0        []\n",
      "group_conv2d_2                 (None, 4, 4, 8, 12)      10368    [(3, 3, 96, 12)]\n",
      "max_pooling3d                  (None, 4, 4, 1, 12)      0        []\n",
      "lambda                         (None, 4, 4, 12)         0        []\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 4 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 3 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "from layers import GroupConv2D\n",
    "\n",
    "input_shape = (8,8,2)\n",
    "\n",
    "x = x_in = Input(input_shape)\n",
    "x = GroupConv2D(12, 3, group='D4', padding='same', activation='relu')(x)\n",
    "x = BatchNormalization()(x)\n",
    "x = GroupConv2D(12, 3, group='D4', padding='same', activation='relu')(x)\n",
    "x = AvgPool3D(pool_size=(2,2,1), strides=(2,2,1), padding='same')(x)\n",
    "x = GroupConv2D(12, 3, group='D4', padding='same', activation='relu')(x)\n",
    "x = MaxPool3D(pool_size=(1,1,x.shape[-2]))(x)\n",
    "x = Lambda(lambda x: K.squeeze(x, axis=-2))(x)\n",
    "model = Model(x_in, x)\n",
    "\n",
    "print(model.input_shape)\n",
    "print(model.output_shape)\n",
    "\n",
    "for l in model.layers:\n",
    "    print('%-30s %-24s %-8i'%(l.name, str(l.output_shape), l.count_params()), [tuple(w.shape) for w in l.weights])\n",
    "\n",
    "d = np.random.randn(*input_shape)\n",
    "\n",
    "# check rotation equivariance\n",
    "d_rot = np.stack([np.rot90(d, k=k) for k in range(4)])\n",
    "y_pred = model(d_rot)\n",
    "y_pred_inv = [np.rot90(y_pred[k], k=-k) for k in range(4)]\n",
    "for k in range(4):\n",
    "    plt.subplot(141+k)\n",
    "    y = y_pred_inv[k][:,:,1]\n",
    "    plt.imshow(y, vmin=np.min(y), vmax=np.max(y))\n",
    "    if k > 0:\n",
    "        plt.title('%.5f' % np.mean(np.abs(y_pred_inv[0]-y_pred_inv[k])))\n",
    "plt.show()\n",
    "\n",
    "# check reflection equivariance\n",
    "d_flip = np.stack([d, np.flip(d,0), np.flip(d,1)])\n",
    "y_pred = model(d_flip)\n",
    "y_pred_inv = [y_pred[0], np.flip(y_pred[1],0), np.flip(y_pred[2],1)]\n",
    "for k in range(3):\n",
    "    plt.subplot(141+k)\n",
    "    y = y_pred_inv[k][:,:,1]\n",
    "    plt.imshow(y, vmin=np.min(y), vmax=np.max(y))\n",
    "    if k > 0:\n",
    "        plt.title('%.5f' % np.mean(np.abs(y_pred_inv[0]-y_pred_inv[k])))\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x144 with 8 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "def plot_indices(indices):\n",
    "    assert len(indices.shape) == 4\n",
    "    nto, nti, k, k = indices.shape\n",
    "    for m in range(nti):\n",
    "        plt.figure(figsize=(nto,2))\n",
    "        for l in range(nto):\n",
    "            plt.subplot(1,nto, l+1)\n",
    "            idxs = indices[l,m,:,:]\n",
    "            plt.imshow(idxs)\n",
    "            #plt.imshow(idxs, vmin=0, vmax=np.max(indices))\n",
    "            plt.xticks([]); plt.yticks([])\n",
    "            for j in range(idxs.shape[0]):\n",
    "                for i in range(idxs.shape[1]):\n",
    "                    plt.text(i,j,idxs[j,i], ha='center',va='center')\n",
    "        plt.show()\n",
    "\n",
    "plot_indices(model.layers[1].indices)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Convert the GroupConv2D model into a normal Conv2D model and get same performance during inference."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x144 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "from layers import GroupConv2D, Conv2D\n",
    "\n",
    "input_shape = (8,8,2)\n",
    "\n",
    "d = np.random.randn(1, *input_shape)\n",
    "\n",
    "x = x_in = Input(input_shape)\n",
    "x = GroupConv2D(12, 3, group='D4', padding='same', activation='relu')(x)\n",
    "x = GroupConv2D(12, 3, group='D4', padding='same', activation='relu')(x)\n",
    "s = x.shape\n",
    "x = Reshape((s[1], s[2], -1))(x)\n",
    "model_1 = Model(x_in, x)\n",
    "\n",
    "y_pred_1 = model_1(d)\n",
    "\n",
    "x = x_in = Input(input_shape)\n",
    "x = Conv2D(12*8, 3, padding='same', activation='relu')(x)\n",
    "x = Conv2D(12*8, 3, padding='same', activation='relu')(x)\n",
    "model_2 = Model(x_in, x)\n",
    "\n",
    "for i in [1,2]:\n",
    "    model_2.layers[i].set_weights([model_1.layers[i].transformed_kernel])\n",
    "\n",
    "y_pred_2 = model_2(d)\n",
    "\n",
    "plt.figure(figsize=(8,2))\n",
    "plt.subplot(141); plt.title('GroupConv2D')\n",
    "plt.imshow(y_pred_1[0,:,:,0])\n",
    "plt.subplot(142); plt.title('Conv2D')\n",
    "plt.imshow(y_pred_2[0,:,:,0])\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### MaxPoolingWithArgmax2D and MaxUnpooling2D"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(None, 32, 32, 10)\n",
      "[(None, None, None, None), (None, 16, 16, 10), (None, 16, 16, 10)]\n",
      "[(4, 32, 32, 10), (4, 16, 16, 10), (4, 16, 16, 10)]\n",
      "\n",
      "[[0.4659 0.2935]\n",
      " [0.0227 0.4706]]\n",
      "0.47061765 330\n",
      "[[0.     0.    ]\n",
      " [0.     0.4706]]\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1008x216 with 4 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "from layers import MaxPoolingWithArgmax2D, MaxUnpooling2D\n",
    "\n",
    "size = 2\n",
    "x = x_in = Input(shape=(32, 32, 10))\n",
    "x, m = MaxPoolingWithArgmax2D(pool_size=size)(x)\n",
    "x_p = x\n",
    "x = MaxUnpooling2D(size=size)([x,m])\n",
    "model = Model(x_in, [x,m,x_p])\n",
    "\n",
    "print(model.input_shape)\n",
    "print(model.output_shape)\n",
    "\n",
    "batch = np.random.rand(4,32,32,10).astype('float32')\n",
    "\n",
    "preds = model.predict(batch)\n",
    "print([e.shape for e in preds])\n",
    "\n",
    "for i in [0]:\n",
    "    print()\n",
    "    print(batch[0][:size,:size,i])\n",
    "    print(preds[2][0][0,0,i], int(preds[1][0][0,0,i]))\n",
    "    print(preds[0][0][:size,:size,i])\n",
    "    \n",
    "plt.figure(figsize=(14,3))\n",
    "plt.subplot(141); plt.title('input')\n",
    "plt.imshow(batch[0,:,:,0])\n",
    "plt.subplot(142); plt.title('maxpooling')\n",
    "plt.imshow(preds[2][0,:,:,0])\n",
    "plt.subplot(143); plt.title('indices')\n",
    "plt.imshow(preds[1][0,:,:,0])\n",
    "plt.subplot(144); plt.title('unpooling')\n",
    "plt.imshow(preds[0][0,:,:,0])\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### MaxPoolingWithRelativePosition2D"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(None, 24, 32, 10)\n",
      "[(None, 24, 32, 10), (None, 24, 32, 20)]\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1224x165.6 with 6 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "from layers import MaxPoolingWithRelativePosition2D\n",
    "\n",
    "input_shape = (24, 32, 10)\n",
    "x = x_in = Input(shape=input_shape)\n",
    "x, p = MaxPoolingWithRelativePosition2D(pool_size=11)(x)\n",
    "model = Model(x_in, [x,p])\n",
    "\n",
    "print(model.input_shape)\n",
    "print(model.output_shape)\n",
    "\n",
    "batch = np.random.rand(4,*input_shape).astype('float32')\n",
    "\n",
    "preds = model(batch)\n",
    "\n",
    "plt.figure(figsize=(17,2.3))\n",
    "plt.subplot(141); plt.title('input')\n",
    "plt.imshow(batch[0,:,:,0])\n",
    "plt.subplot(142); plt.title('max pooling')\n",
    "plt.imshow(preds[0][0,:,:,0])\n",
    "plt.subplot(143); plt.title('relative position in y')\n",
    "plt.imshow(preds[1][0,:,:,0]); plt.colorbar()\n",
    "plt.subplot(144); plt.title('relative position in x')\n",
    "plt.imshow(preds[1][0,:,:,input_shape[-1]]); plt.colorbar()\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### AddCoord2D"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(None, 28, 28, 3)\n",
      "(None, 28, 28, 5)\n",
      "[[ 0.4147  0.2122  0.4213 -1.      0.9259]\n",
      " [ 0.277   0.7857  0.39   -0.9259  0.9259]\n",
      " [ 0.1146  0.0171  0.1901 -0.8519  0.9259]\n",
      " [ 0.9058  0.5574  0.0226 -0.7778  0.9259]\n",
      " [ 0.2437  0.0084  0.559  -0.7037  0.9259]\n",
      " [ 0.041   0.4743  0.4853 -0.6296  0.9259]\n",
      " [ 0.4955  0.8261  0.0195 -0.5556  0.9259]\n",
      " [ 0.0021  0.3706  0.508  -0.4815  0.9259]\n",
      " [ 0.556   0.6704  0.9399 -0.4074  0.9259]\n",
      " [ 0.331   0.3188  0.5686 -0.3333  0.9259]\n",
      " [ 0.9825  0.283   0.1884 -0.2593  0.9259]\n",
      " [ 0.5171  0.5946  0.0987 -0.1852  0.9259]\n",
      " [ 0.4246  0.4799  0.4373 -0.1111  0.9259]\n",
      " [ 0.8     0.7057  0.7673 -0.037   0.9259]\n",
      " [ 0.0434  0.3547  0.2791  0.037   0.9259]\n",
      " [ 0.5431  0.6299  0.1728  0.1111  0.9259]\n",
      " [ 0.2309  0.578   0.4045  0.1852  0.9259]\n",
      " [ 0.5644  0.0977  0.7081  0.2593  0.9259]\n",
      " [ 0.3113  0.6464  0.8978  0.3333  0.9259]\n",
      " [ 0.1136  0.4867  0.8515  0.4074  0.9259]\n",
      " [ 0.8031  0.8038  0.6576  0.4815  0.9259]\n",
      " [ 0.3704  0.3949  0.9279  0.5556  0.9259]\n",
      " [ 0.6792  0.339   0.6366  0.6296  0.9259]\n",
      " [ 0.4639  0.5134  0.7409  0.7037  0.9259]\n",
      " [ 0.4403  0.1509  0.1581  0.7778  0.9259]\n",
      " [ 0.6374  0.4545  0.7302  0.8519  0.9259]\n",
      " [ 0.695   0.2277  0.1998  0.9259  0.9259]\n",
      " [ 0.992   0.6593  0.1006  1.      0.9259]]\n"
     ]
    }
   ],
   "source": [
    "from layers import AddCoords2D\n",
    "\n",
    "x = x_in = Input(shape=(28, 28, 3))\n",
    "x = AddCoords2D()(x)\n",
    "model = Model(x_in, x)\n",
    "\n",
    "print(model.input_shape)\n",
    "print(model.output_shape)\n",
    "\n",
    "batch = np.random.rand(4,28,28,3).astype('float32')\n",
    "\n",
    "preds = model.predict(batch)\n",
    "print(preds[0][-2,:,:])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### DepthwiseConv2D"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(None, 28, 28, 3)\n",
      "(None, 28, 28, 32)\n",
      "[ 0.031  -0.2327 -0.3472 -0.3636 -0.04   -0.2751 -0.1038  0.1884 -0.3158 -0.005   0.3535 -0.0798 -0.1447\n",
      "  0.019  -0.5067 -0.1077  0.1027 -0.0345  0.1869  0.2644 -0.2179  0.3228  0.2713 -0.2901  0.0245  0.0948\n",
      " -0.0273 -0.0285  0.3437  0.4658 -0.0039 -0.3469]\n"
     ]
    }
   ],
   "source": [
    "from layers import DepthwiseConv2D\n",
    "\n",
    "cm = 2 # channel_multiplier\n",
    "\n",
    "x = x_in = Input(shape=(28, 28, 3))\n",
    "x = Conv2D(16, (3,3), padding='same')(x)\n",
    "x = DepthwiseConv2D(cm, (3,3), padding='same')(x)\n",
    "x = Conv2D(16*cm, (1,1), padding='same')(x)\n",
    "model = Model(x_in, x)\n",
    "\n",
    "print(model.input_shape)\n",
    "print(model.output_shape)\n",
    "\n",
    "batch = np.random.rand(4,28,28,3).astype('float32')\n",
    "\n",
    "preds = model.predict(batch)\n",
    "print(preds[0][0,0,:])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### LayerNormalization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(None, 28, 28, 3)\n",
      "(None, 28, 28, 3)\n",
      "0.08324947 0.99998516\n"
     ]
    }
   ],
   "source": [
    "from layers import LayerNormalization\n",
    "\n",
    "x = x_in = Input(shape=(28, 28, 3))\n",
    "x = LayerNormalization()(x)\n",
    "model = Model(x_in, x)\n",
    "\n",
    "print(model.input_shape)\n",
    "print(model.output_shape)\n",
    "\n",
    "batch = np.random.rand(4,28,28,3).astype('float32')\n",
    "\n",
    "preds = model.predict(batch)\n",
    "\n",
    "print(np.var(batch), np.var(preds))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Conv2D (Weight Normalization)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(None, 28, 28, 3)\n",
      "[(None, 26, 26, 10), (None, 26, 26, 10)]\n",
      "0.084439 8.107712 0.21911137\n"
     ]
    }
   ],
   "source": [
    "from layers import Conv2D, gaussian_init\n",
    "\n",
    "x = x_in = Input(shape=(28, 28, 3))\n",
    "x1 = Conv2D(10, 3, kernel_initializer=gaussian_init)(x)\n",
    "x2 = Conv2D(10, 3, kernel_initializer=gaussian_init, weightnorm=True)(x)\n",
    "model = Model([x_in], [x1, x2])\n",
    "\n",
    "print(model.input_shape)\n",
    "print(model.output_shape)\n",
    "\n",
    "batch = np.random.rand(4,28,28,3).astype('float32')\n",
    "preds = model.predict(batch)\n",
    "\n",
    "print(np.var(batch), np.var(preds[0]), np.var(preds[1]))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Resize2D"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(None, 12, 12, 1)\n",
      "[(None, 6, 6, 1), (None, 24, 24, 1), (None, 6, 6, 1), (None, 24, 24, 1)]\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1152x216 with 5 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "from layers import Resize2D\n",
    "\n",
    "s = 12\n",
    "s1, s2 = s//2, s*2\n",
    "\n",
    "x = x_in = Input(shape=(s, s, 1))\n",
    "x1 = Resize2D((s1,s1), 'bilinear')(x)\n",
    "x2 = Resize2D((s2,s2), 'bilinear')(x)\n",
    "x3 = Resize2D((s1,s1), 'nearest')(x)\n",
    "x4 = Resize2D((s2,s2), 'nearest')(x)\n",
    "model = Model(x_in, [x1,x2,x3,x4])\n",
    "\n",
    "print(model.input_shape)\n",
    "print(model.output_shape)\n",
    "\n",
    "batch = np.random.rand(4,s,s,1).astype('float32')\n",
    "\n",
    "preds = model.predict(batch)\n",
    "\n",
    "plt.figure(figsize=(16,3))\n",
    "plt.subplot(151); plt.title('input')\n",
    "plt.imshow(batch[0,:,:,0])\n",
    "plt.subplot(152); plt.title('bilinear %.1f' % (s1/s,))\n",
    "plt.imshow(preds[0][0,:,:,0])\n",
    "plt.subplot(153); plt.title('bilinear %.1f' % (s2/s,))\n",
    "plt.imshow(preds[1][0,:,:,0])\n",
    "plt.subplot(154); plt.title('nearest %.1f' % (s1/s,))\n",
    "plt.imshow(preds[2][0,:,:,0])\n",
    "plt.subplot(155); plt.title('nearest %.1f' % (s2/s,))\n",
    "plt.imshow(preds[3][0,:,:,0])\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Blur2D"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAA6AAAADJCAYAAAA91l++AAAABHNCSVQICAgIfAhkiAAAAAlwSFlzAAALEgAACxIB0t1+/AAAADh0RVh0U29mdHdhcmUAbWF0cGxvdGxpYiB2ZXJzaW9uMy4xLjIsIGh0dHA6Ly9tYXRwbG90bGliLm9yZy8li6FKAAAgAElEQVR4nO3de5Rc1X0n+u+3q6tb6lbrLR56GSNhbEhs2VFAJJYGP8GsTAh3PINxxiYaZwmy8FybcG9EnHttx8lkmdyxde8MxAIHEM74gWOjmMlgY+yLb4uJxMNEvAQOLSIDktATodajn/W7f9QR6a6zu/dP3VWnqk5/P2v16q5du/b5nVO/qq5d55zfoZlBREREREREpNZa6h2AiIiIiIiITA2agIqIiIiIiEgmNAEVERERERGRTGgCKiIiIiIiIpnQBFREREREREQyoQmoiIiIiIiIZEIT0IyQfI7kpfWOQ0REROQUkrtIfnCM+y4l+WrWMYlUi/K7MWkCmhEzu9DMflbLZZDcRPLPa7kMkXojeSPJl0geJbmH5AaSrfWOS2QySP4lyVeSvP4lyc/VOyaRaks+8JdIHhvxc2294xKpBpKfq8jtk0m+z693bI1GE1ARaTb3A3iPmc0E8CsA3gXgf61vSCKTdieAtyd5/RsAfpfk/1LnmEQmbJwvBveY2YwRP/dkGphIFYTy28z+YmRuA7gFwM/M7GD2ETY2TUAzcuoQAJJfJPldkt8g2Zscmruyot8fk9xB8nWSd5Ocltz3eyQfqRjXSC4nuQ7A7wL4o+Rbl/+e7RrKVJPk6v9O8mmSx0neSfJMkj9McvsnJOeM6P+3JF8j+QbJbpIXJu1tJLeT/I/J7QLJ/0ny86HlmtlOMztyalgAJQDLa7y6MkWRvJnkziSnd5C8KmlvJ3mE5K+M6Lsg+cb7jOT2H5Hcm+yp//1T79eh5ZjZL8zs+Igm5bVk6ddDnzsqVebwyCOvTh3OSHI9ydcA3J1R7CIxmec3SQL4JAB9wRKgCWh9/DaA7wCYjfLenFsr7v9dAJcBWAbgbQD+j9iAZnYHgG8C+Mvkm5d/XdWIRcL+DYAPoZyn/xrADwF8DsAClN9fRu6Z/CGA8wCcAeBJlPMVZjYA4N8D+BLJdwC4GUABwH8aa6EkP07yKICDKO8Bvb2qayXyL3YCWA1gFoA/BfDfSJ5tZv0A7gNwzYi+/w7A/2dm+0leDuAPAXwQ5YnkpbEFJZPdYwBeBdAJ4FvVXBGRcZz2544xnAVgLoC3AFg3Rp8zSO4j+c/JKRSdE1yWiFeW+X3KapQ/73x/gsvKNU1A6+MRM3vAzIYB/A3KH6BHutXMXjGzwyh/CL8mNYJIY/ivZrbPzHYD2ALgUTP7RzPrA7AZwLtPdTSzu8ysN/ng/kUA7yI5K7nvWQB/DuDvAPxvAD6RvD6CzOxbyaGKbwOwEcC+2qyeTHVm9rdmtsfMSmZ2L4AXAVyU3P0tAB8b0f3j+JdJ478DcLeZPWdmJ1DO+diyvgygC8B7UP7f8EZ11kIkqlqfO0oAvmBm/WZ2MnD/CwBWADgbwPsB/BqAr05wWSJeWeX3SNcC+J6ZHZvgsnJNE9D6eG3E3ycATKs4lvyVEX//EsDCTKISOX0jJ34nA7dnAG8eVvvl5FDGowB2JX1Gnph/D8rfKj5gZi96Fp70ew7AX00sfJHxkfxkcoj4EZJHUD7v+FTePgygg+TFJM9B+YP15uS+hRj9Xj7y7zFZ2T+i/Pr50yqsgohHtT53HEi+gAwys9fMbEfyhc4/A/gjlI+kEamlTPL7FJIdAP4tdPjtmDQBbUxLRvy9FMCe5O/jADpO3UHyrIrHWY3jEpmojwO4EuXDEWcBOCdp54g+fwXg7wFcRvK9pzF2K8qH1YhUFcm3APg6gE8DmGdmswE8iyRvk73030X52/RrAPy9mfUmD98LYPGI4Ua+r3soryVLY33uqHQCIz6HoHxI4kin+znEoM+iUntZ5/dVAA4D+Jmz/5SjF31juoHkYpJzAfwJgHuT9qcAXEhyRXIC9RcrHrcPwLnZhSni1gWgH8AhlN/c/2LknSQ/gfKhWL+H8nmj95CcERooKeZyqsjLBQD+GMBPaxa5TGWdKH/gOAAAJNeivAd0pG8BuBrlc4xGnrP5XQBrSb4j+Tb8/xxrISRbSF5Hcg7LLgJwA5TXkp2xPndU2g7g48lRLZcD+FensxCS7yP5liTPlwD4MoAfTCpykbhM8nuEawF8w8y0Y2gMmoA2pm8B+DGAl1AugPHnAGBm/wTgSwB+gvJ5SI9UPO5OABckh4r9XXbhikR9A+XDXnYD2AFg26k7SC4F8H8D+KSZHTOzbwF4AsCGMcb6TQDPkDwO4IHkR9dMlKozsx0AvgJgK8pf8P0qgP9Z0edRlI9OWYhyoa1T7T8E8F9QPky3B/+S8/1jLO4qlN/vewH8NwD/NfkRyULwc0fAZ1AuOHcE5S9dTvezxrsB/APKr5l/APAMdBktqb2s8hskF6F8fvM3JhTpFEFNzhsLyV0Aft/MflLvWEREpDqSCs/PAmg3s6F6xyMiIlIv2gMqIiJSAySvSq4XOgflC5L/d00+RURkqtMEVEREpDauA7Af5UO+hgH8QX3DERERqT8dgisiIiIiIiKZ0B5QERERERERycSkJqAkLyf5C5I9JG+uVlAiIiIiIiKSPxM+BJdkAcA/AfgQgFcBPA7gmqRsfVBhRqe1zp07qq3Ym+739qUHRt3eNZC+HGDv8emuONtfOZ5qa3v76Hl332vpsaaddTLVdk7bsVRbKLbQeC1HRsfRv6Qz1aerM73MgRdKqTYZXy9eP2hmC+qx7PlzC3bOkmJVxnrmdd8q/OqcA/FOVRbK+4lSjp++euV4obPTirPnRvtZIT5Wy4BvmaU2R6fW6uZQy8nqHRxUfC39P0jGV8/3cHeOt8Y/O3GIrmV6Xi9gdU+X8sRG58uq7UCfq58ND/sGnALqnuNzHDnueBv05ohnLHeOm+91RUe60VkSru31sa6eNZoNOgb0hV++AnUTi+V46yTGvghAj5m9BAAkvwPgSpSv8Rde2Ny5WHjTZ0e1LexOb+Ett90+6vbal1en+nRvvdAV5PIbt6XaFt7TNep2zy0XpB+3Pr0ady/dkmoLxRYar2Pzo6P73LQq1WfNJc+l2vasCszQZVw/se/90tMvucjw/wOgAOCvzezLFfe3o3wdp18DcAjA1Wa2a7wxz1lSxGMPLplI2CnL7r3e1e+xqzdWZXmnI5T3E6UcP32eHK9Ffhdnz8XS6/8wGt/gzPinko7dvkneiUWOTzjzfR8OvKY97/uC02PxX/xD1caaKur5Hl6cPRdL/+DG6LL758c/3bYf9MwsgYE58Ry3YnU/jU57Lf7xz/sl0Vv+Kv3ZJWT4yBu+AaeAuub4nLlY/Ol4jg/NiOdl6zHf+/hwRzx/S9N8s1kO+mZwba/HY2s/5Btr8XdfcvUbem1ftA8LvvcFKzle86XG/VInluOT+Zp3EYBXRtx+NWkbheQ6kk+QfGL4mL4JlsaQ7MG/DcBHAFwA4BqSld8cfArA62a2HMAGlC+jINLwlN+Sd8pxyTvluORZzYsQmdkdZrbSzFYWZqQPOxWpkzf34JvZAIBTe/BHuhLAPcnf3wPwAZLegydE6kn5LXmnHJe8U45Lbk1mArobwMhjDRcnbSLNwLMH/80+ycXj3wAwL5PoRCZH+S15pxyXvFOOS25N5hzQxwGcR/KtKE88Pwbg4+M9oNibPudzz5r0FzWXLVwx6vaDe9LnXi5zngN64qqLU23dW0cvc81637mXa7f5zkUNjdeD0XGEzveU5kRyHYB1ALB00WReUiKNZ2R+t86aU+doRKpPOS55NyrHZyvHpf4mvAc0+abl0wAeBPA8gO+amWZV0iw8e/Df7EOyFcAslE/yH2XkYeYL5vlOLhepsZrkd6FTp1FIw1COS94pxyW3JnUOqJk9YGZvM7NlZvafqhWUSAbe3INPsg3lPfj3V/S5H8C1yd8fBfD/2kSvWySSLeW35J1yXPJOOS65peMFZUoysyGSp/bgFwDcZWbPkfwSgCfM7H4AdwL4G5I9AA6j/OYv0vCU35J3ynHJO+W45JkmoDJlmdkDAB6oaPv8iL/7APzbrOMSqQblt+SdclzyTjkueZXpBHSwK110KFSMp6eicNDqG9KFhBYifYRBqKBRqG3n1RtH3V527/WpPmu2peMKFRyqLKoEAN0IFEhaU3H7lspLOQHL1+9IP06ayq6BGVj7crpYVb01Ykyn68E92zNdXmUxNAGs1dC/YCjar6U/fnZHqc23zFJ7/OLkLQfbXWMV+n1XJxicGT+Cbf726h7lVnjbsmiffZee4Rpr5svx56jtR4+7xppqrOB7/lmK51LJ+wnLkZbFw776Ai2DvhwvtcbXcdaLvhy3vn5Xv9Zzz4n2OfQbZ7nGmn4gnuPtP3vGNZb1++LPC2sBhjscz60jlayKF3NsPeLM8WHngBZfgY4D8f8vAGADA65+hbcvj/Y59OvzXWN5cnz6Iy+4xir1pgur1lvNrwMqIiIiIiIiAmgCKiIiIiIiIhnRBFREREREREQyUfciRD2OcyFD5156hc7RXHvJ6mifnu50XKnzOAF0bH403W19V3q8ivUMnZsaOi+0A+nxRUREREREmpH2gIqIiIiIiEgmNAEVERERERGRTGgCKiIiIiIiIpnQBFREREREREQyUfciRKFiPJWXcV1zyXOpPpMpTFT52DXr0+OHiiMF49iwKv3YrYGFBgoYZe3BPdurNtZlC1dUbSwJC+Vbo/C8/ho5fqmx+fELuw/2T6va4opHfd+lltodF18H0H4wfgHzQr/vAuZe+y49I9rn51/8mmusVds/Gu3T9iPXUFMPDaVpjue2fTjapXSyzbfMYUe+DcT7AECp1ZfjxWPx8VqrnOOHfuOsaJ/PfeEbrrG+sOO3o33OfnKma6zhAwdc/XKDhlK747ltieeSOfPSHGMV+n3v44y/9AAArX2OPid9rxfQt56Hfn1+tM+v/cd/dI31ox2BYqgV3vHCXNdYpd5eV78saQ+oiIiIiIiIZEITUBEREREREcmEJqAiIiIiIiKSiUmdA0pyF4BeAMMAhsxsZTWCEhERERERkfypRhGi95nZQU/H9leOY/mN20a1LdzWVYUQxrZ8/Y50Y0WBoZ7u9Im+oeJIewKFV6pdIElERERERCSvdAiuTDkkl5B8mOQOks+R/Eygz6Uk3yC5Pfn5fD1iFZkI5bjknXJc8k45Lnk22T2gBuDHJA3A7WZ2R2UHkusArAOAaeiY5OJEqmIIwE1m9iTJLgA/J/mQmVXuLt9iZr9Vh/hEJks5LnmnHJe8U45Lbk12D+h7zew9AD4C4AaSqatdmtkdZrbSzFYW0T7JxYlMnpntNbMnk797ATwPYFF9oxKpHuW45J1yXPJOOS55NqkJqJntTn7vB7AZwEXVCEokKyTPAfBuAI8G7r6E5FMkf0hSJ/ZKU1KOS94pxyXvlOOSNxM+BJdkJ4AWM+tN/v4wgC+N95jS7E6ceP/Fo9q6t6aL/VQW9tmzqjc92IZ008JuS7V1I/1a3HnbxlG3l917fSjclMoCSgDQc9XF6Y6p/cDhx1YKFWTas9kVmkwAyRkAvg/gs2Z2tOLuJwG8xcyOkbwCwN8BOG+Mcd48zLzzrM4aRlx7dy/d4uvo6Lf25dWTjEYmqxo5PjK/C/NmoaVrMLrc4kvTo336FwxF+wDAuW97LdrnyHbfToH2o+n/ESG9i9P/lyr1z/R9f+s98WTmy/HtsWr7R11jHXn8jGifWehxjdXoqp/js8Hp8eei+HL8iK7BuaVoHwCYveRItM/AnnmusWbsd3XDwOx4n/4uX453FQquftMPxLfrF3b8tmusY8/PiXca3Osaq9HVIsdRjOdm6+FitM9wpy/HW+edjPZpOTDDNVb7YVc3DDnefAdmOPfDVTHHf7QjXfg0pH3ntHinvn7XWI1oMntAzwTwCMmnADwG4H+Y2Y+qE5ZIbZEsovyG/k0zu6/yfjM7ambHkr8fAFAkOT801sjDzKfNdrxhiGSgWjk+Mr8LM5v7CxbJl5rkeJdyXBpHTXJ8hnJc6m/Ce0DN7CUA76piLCKZIEkAdwJ43sy+OkafswDsMzMjeRHKX9YcyjBMkQlTjkveKccl75TjkmfVuA6oSLP5TQCfAPAMye1J2+cALAUAM9sI4KMA/oDkEICTAD5mZr7j90TqTzkueaccl7xTjktuaQIqU46ZPQJg3JO8zOxWALdmE5FIdSnHJe+U45J3ynHJs0wnoINdwJ41o19LlQWHQk4ECv2ECg5Vjj3W+LUujhKKrXIdQrH2bE2PtRzx4kUiIiIiIiLNYLLXARURERERERFx0QRUREREREREMqEJqIiIiIiIiGQi03NAuzpPps7J7N56YfyBa3zje84n9T6u5xbfhWJDttx2e6pt9Q3X1XSZHpctXFHT8aVx3b10S8Mu8zL48lL52wCGiVJv/OLk/QviF+NGm+8C5i/tOiPe6aJB11heXfOPR/scfWqOayxfL6DtR487+vjGmoUe51IlxQDri190fqgrXmjUCr5ipK/vmRXvtNyX4yeWOPcrzIi/Rtt3tbuGWlD0fZRs/9kz0T5nPznTNRYG90a7DL9x1DfWVGMABuN5MtzuyPEWX44PvhHPpcGzHf83APQtGLcu05tsevx/jLXE/58BwPwW3+tq+iMvRPu844W5rrHQ1x/tMnzgoG+sBqQ9oCIiIiIiIpIJTUBFREREREQkE5qAioiIiIiISCY0ARUREREREZFMZFqEqPf49FTRIU/hoFBxnj1rfCchewr7hIoGrV0f6Li+K9XUvdUXR/px6eJLO2/bmGq7bLMKr4iIiIiISD5oD6iIiIiIiIhkQhNQERERERERyYQmoCIiIiIiIpKJ6ASU5F0k95N8dkTbXJIPkXwx+e291raIiIiIiIhMUZ4iRJsA3ArgGyPabgbwUzP7Msmbk9uhsj2jtL9yHMtv3DaqrXvDqngEawJNgeJFEy1WtPqG61yPCy1zYbel2pbh+vRj149+7PJAHGtfXh1o7Q20SaM6p+0Y7l66Jdov/FxPjGcsT0zVXqafcrxZcJgoHi1E+w23p98XK5XaSq5ldrzUFu1zYvGQa6y2eX2ufv39xWifiZWfk4Y3RLS+Hv9o5Hn+S/HUBQC0H4y/pvrP9OV4qzPHQcdrtOBcgVJ8LACw/v5on+EDB3zLlIkbJlod7+OuoQq+d8KCY3lDc3w5Xpg94OrnycpSW/y9/nSUeuOfZzx9poLoHlAz6wZwuKL5SgD3JH/fA+B3qhyXSM2R3EXyGZLbST4RuJ8k/wvJHpJPk3xPPeIUmQjlt+SdclzyTjkueTXRy7CcaWZ7k79fA3DmWB1JrgOwDgCmoWOCixOpmfeZ2cEx7vsIgPOSn4sBfC35LdIslN+Sd8pxyTvluOTOpIsQmZlhnD3dZnaHma00s5VFtE92cSJZuhLAN6xsG4DZJM+ud1AiVaL8lrxTjkveKcelKU10ArrvVIInv/dXLySRzBiAH5P8ebKnvtIiAK+MuP1q0ibSDJTfknfKcck75bjk0kQPwb0fwLUAvpz8/oHnQf1LOtFzU7zoUGVhn1BBoO6tF6Yf6CxWVPnYjs2Pph+3visS5dixeQTjD1iObfFOMlHvNbPdJM8A8BDJF5Jznk/LyMPMly6a6EtKpOqqnt+ts1XwXBqKclzyTjkuueS5DMu3AWwFcD7JV0l+CuWJ54dIvgjgg8ltkaZiZruT3/sBbAZwUUWX3QCWjLi9OGmrHOfNw8wXzKtOZTmRyapFfhc6O2sVrshpq0WOtyjHpYEoxyWvPFVwrzGzs82saGaLzexOMztkZh8ws/PM7INmVlklV6Shkewk2XXqbwAfBvBsRbf7AXwyqTK3CsAbI4pviTQs5bfknXJc8k45Lnmm4wVlqjoTwGaSQPl18C0z+xHJ6wHAzDYCeADAFQB6AJwAsLZOsYqcLuW35J1yXPJOOS65pQmoTElm9hKAdwXaN4742wDckGVcItWg/Ja8U45L3inHJc8ynYD+6pwDeOzqjaPalt17fapfZWGfyqJEALB8/Y5UW6iwT6itcryeDenCSCduSS9zy223p9pCXzV54ggVRwrZ4+olU9ndS7fUOwSZIqwFGG4f86pbbyoejRdYt4UDrmX2vWM42qetLd4HAIYGfOdol/rj/XRRsamtcDJegLC0ZNA11mBXvF+xteQaa3jYd3EDG4r3K7pGAtAysWKMUh8EQN9bZtyMIVe34a74/41Ca7wPAAw738dt2JGXvpcVUHDW92jJuA6IeVegiui8gEokxyZ9HVARERERERERD01ARUREREREJBOagIqIiIiIiEgmNAEVERERERGRTGRahOiFlxdg9Q3XjW5cM7Gxem65INW2Zr2vsA8uGX1zT6BoUKjI0dqXV7uG9xRNChUq8hYmEhERERERaUbaAyoiIiIiIiKZ0ARUREREREREMqEJqIiIiIiIiGQi03NAW44cR8fmR0e1rVnfleoXOj+yUugczRDPWDuv3phqW3bv9am20Lmde9YELnQbOK91eTQKyYNdAzPc5wo3olDeh1TzfOWF29LvASF3L91StWV6eLeFV1XP8b64ekOdjpZBYNqB+PeWfQscF8d2Xky8s6sv2qf3YKdrrBkvtLn69c93XDS93zUUWt71Dle/vrNnxJfZV62rxwPF109WbSyvY+fO9HW873u1DWQcNKC1L34B+8GZ8Rz3XiO+2D4U7dN/3Je703e2u/oNzYjneMtQfDsAwNAF57j6DU+r3kdOWjz+4uvx9w4A4KDzdVWKP6F9S2b5xnqwfjkOA+hY5aHO+DYGHX0AtDpyfPCEL8fbdhdd/Uqel4JzN9yJC8929Ws5/6xoHyv4XlccduT40QHfWP2Drn5gPLb+M33/b/HgvePerT2gIiIiIiIikglNQEVERERERCQTmoCKiIiIiIhIJqITUJJ3kdxP8tkRbV8kuZvk9uTnitqGKSIiIiIiIs3Oc0b4JgC3AvhGRfsGM/vPp7Ow/iWd6Llp1ai2E7cETrKtKOITKvSzJ1BcKFTkI9TWc8sFoxuuDgQbECw4FBAqVtTTXbHMQKGiUMGk5djmWqaIiIiIiEiji+4BNbNuAIcziEUkEyTPH7H3fjvJoyQ/W9HnUpJvjOjz+XrFK3K6lOOSd8pxyTvluOTZZGpif5rkJwE8AeAmM3s91InkOgDrAKAwZ84kFidSHWb2CwArAIBkAcBuAJsDXbeY2W9lGZtINSjHJe+U45J3ynHJs4kWIfoagGUovzD2AvjKWB3N7A4zW2lmKwsznNeOEcnOBwDsNLNf1jsQkRpRjkveKccl75TjkisTmoCa2T4zGzazEoCvA7ioumGJZOZjAL49xn2XkHyK5A9Jpk/QFWkOynHJO+W45J1yXHJlQofgkjzbzPYmN68C8Ox4/U8p9qYL9HgL+3ikiguNMf7O2zaOur36huvSgwWKBIWECg6FVMYRKo4UfFygrWfDqlRb5Xh3L92S6hNazy233Z5qW3bv9am2nVfHt9lknkvPdly+fodvsIt93Ui2AfhtAH8cuPtJAG8xs2NJlee/A3DeGOO8eZj50kWtwW1fae3Lq31BZsybl551bHbebdHIqpHjo06jmDsbJxYPVSU277efvQerd+TMYJez38zhaJ9pb+91jfXCud744++BxYPtrpE88bcfqO4pMdMOxvu8caEzd+7zL7fqOT5vNvoWD8QXPODI4CFflvefnBbt0zLdt+0GZ/g+k5SmxfsNL+pzjbVzznRXP5TiXTjs+xwx3Ol4je6f7Rqr9birG9pfj2+zN97mGwsPOvuhNu/jA/Pi248Wfy7MmeODg23xTs5/CsPxlwsAoNQeT7ihufHtAACvzPZNlzgU32Z0vA4AoFR0/E846vuf0HbUt8zi8fgyT5zp/KwfyXHPZVi+DWArgPNJvkryUwD+kuQzJJ8G8D4AN/qiEWkoHwHwpJntq7zDzI6a2bHk7wcAFEnODw0y8jDzBfMKtY1Y5PRMOsd1GoU0OOW45J1yXHInOqU3s2sCzXfWIBaRrF2DMQ5pIXkWgH1mZiQvQvnLmkNZBidSBcpxyTvluOSdclxyZzJVcEWaFslOAB8CcN2ItusBwMw2AvgogD8gOQTgJICPmZnv2CaRBqAcl7xTjkveKcclrzQBlSnJzI4DmFfRtnHE37cCuDXruESqRTkueaccl7xTjkteZToBHexKF6qpLG4DhIvgVAoVCOlGuvhXqF9lAZ1goSJnXKHHhgrqVLbdfXW6iEtwvTekm0IqCzCthq8gU6iY0Jr16W1WWTSnY/Oj6SDWpIsjhXgLN1Uuozswfh4KxYiIiIiITBUTvQ6oiIiIiIiIyGnRBFREREREREQyoQmoiIiIiIiIZEITUBEREREREclE3avgVha3AdJFakLFcybDM14orhBvMaSJju8thrTlttujfUJC26JjVW+qbeG2rnFvA0DPVtcisXz9jlRb99bANqsoOhQqXtTTnS62lBd3L00Xqpoob741u2Zfz+DroIFwiGg/EP+3UWqPFxrjvD7XMoccfUq9RddY/Qs8owFtztg8zj1nv6vfrmcWRvu0H/T9Lxyc6ehz7knXWKX+gqtfob8t2mfuk76xXnb1qg0OEW374vlUcqTc0PRh3zIH4/sC7GS7a6yheb4cb+0cjPZhS8k1VtvS9GeGkJMvpz83VGrt9eV4aXq8X99Zvm3RctK5L4bxfl3/3ARXQCkRrcfir8VSMb4u5ktLYCC+7TyvAwAYnuV7Xgsz4jleLPpeo6WuAVe/wX3T48s86lvPUke8jzfHB2f5ltl+MN6v7ahrqCjtARUREREREZFMaAIqIiIiIiIimdAEVERERERERDKR6TmgvzrnAB6rOKfxsoUrUv32bBh9bH/o3D/PeZZjWX7jttHjh85nvCVwbuEa3/ih80InOr73XE7PuW/B81UD55yFtkdlP886jiW07gvRBOdNiIiIiIjIpGgPqIiIiIiIiGRCE1ARERERERHJhOTuWTsAABgtSURBVCagIiIiIiIikonoBJTkEpIPk9xB8jmSn0na55J8iOSLye85tQ9XREREREREmpWnCNEQgJvM7EmSXQB+TvIhAL8H4Kdm9mWSNwO4GcD68QZ64eUFWH3DdaMbr4oHsHz9jnRjoJDNnjXpCxPfvXRLqm3ZhtGFfXq2podfsz5QZCewzJ5uXxyVdt62MdUWKjjkLfZTWSSostASAOwJPG7hVeniP6F1ShUJuiQwVqBYVPC5Czw2JFQgyeW+iT0sS6G8nKo8BbSmCu/rfVdtwxiTtRr6F8QvfN02ry/aZ+DQNNcyW/rjB+p0HPAdzNO3oOTq194ev4B5R1u8DwC8tOsMV79if/z/RqnNNRQ6Xo3/a+9b4NtmLfP7Xf1OnBu/UPtQZ9E1Vj1ZwTAwL35x+uLseI7jdWeOn4w/98Vjzhzv9F2Yfs6s4/FlFuLbAQAOvjHD1a/UER+v1BffFgDQfqAQ7TM40/d6H57u63fyzHgfYxMcWFgwDM2K50lLhyOXjvte04UT8e3SMuR77r3Pa1dn/DVK+gpgHjvuey3btHhspZO+HCn2xrfH0AzfWKU233oOzPL08z1PMdHIzWyvmT2Z/N0L4HkAiwBcCeCepNs9AH6nKhGJVBHJu0juJ/nsiDbX3nuS1yZ9XiR5bXZRi/govyXvlOOSd8pxmYpO66sakucAeDeARwGcaWZ7k7teAxD8bojkOpJPkHxisP/YJEIVmZBNAC6vaLsZ5b335wH4aXJ7FJJzAXwBwMUALgLwBR1mLg1oE5Tfkm+boByXfNsE5bhMMe4JKMkZAL4P4LNmdnTkfWZmQPhCjmZ2h5mtNLOVxXbfYRoi1WJm3QAOVzR79t5fBuAhMztsZq8DeAjpfxAidaX8lrxTjkveKcdlKnJNQEkWUZ58ftPMTp1ht4/k2cn9ZwPYX5sQRarOs/d+EYBXRtx+NWkTaXTKb8k75bjknXJcci1aqYAkAdwJ4Hkz++qIu+4HcC2ALye/fxAba7DLWaDn6tEFekLFeUJFfFIFjgAsQ7ywT0+ooNGNvem2DenYQ0VD9gSK57jW6er0OnkLtKTi2NblehyQLhIU3B4Vz1toHVOFisYw0eJCwW29Kv087TqNMc3M6D0LfQwk1wFYBwBLF3nqeolko9r5XZg7uypxiVSLclzyTjkueeTZA/qbAD4B4P0ktyc/V6A88fwQyRcBfDC5LdIMPHvvdwNYMuL24qQtZeRh5gvmxSvzidRYzfK7MKOz6sGKTIByXPKudjnepRyX+vNUwX3EzGhm7zSzFcnPA2Z2yMw+YGbnmdkHzazy+HWRRnVq7z0w9t77BwF8mOSc5KT+DydtIo1O+S15pxyXvFOOS641wQWLRCaO5LcBbAVwPslXSX4KY+y9J7mS5F8DQPKFyp8BeDz5+ZK+ZJFGo/yWvFOOS94px2Uq0glrkmtmds0Yd30g0PcJAL8/4vZdAO6qUWgik6b8lrxTjkveKcdlKmrICWhl4Z1Q8ZnLFq5ItS3fli6oEyqWEyqyU2lhoIhPz9bow8bkKSYU6uMpCAQAy2/cNup2KP4Q7/iVFnanz4dfvt5X0ChUrMjz2G4EihdtCAT32e8FGrOxa2CGu3BUtdy9dEtV+gD+olfVXEdvbB7euDzLzPp5bAZsNbTN66vKWC391TsAp3DR665+C9oGXf0OHIq/f/bt9hXyaO+Pv58CQP+CoWifwZmuoYC2UrRLS/uwa6jC7mmufkXHeg7OjMdVby3FEjrPOB7tNzwcz9+hAV+Ol4rxPoV3vuEaa2HHSVe/A0fil8UbOjTdNZa1+p7X4qz+aJ/Bku/1Mjgr3ofTfTle3Nvm63c0HlvJN1RdsbWE6fPieTI0GK9pYUO+56s0PV5DqWWuL3dnd8TzCACOHY+/dw0ddT5hBV8NqJaO+Pv4kHObubT64mo97JvuFY/HY7MqlTrRIbgiIiIiIiKSCU1ARUREREREJBOagIqIiIiIiEgmNAEVERERERGRTNS9CFGowJBHz4ZV6TZnkaDKIjuhgjqhIiVrfcO71im0TFziG3/n1RtTbWsvGV0wJRh/oKhKqOBQKP49q3pH3fYWOfIUNAIAOApDhYS2464JjSQiIiIiIrWmPaAiIiIiIiKSCU1ARUREREREJBOagIqIiIiIiEgmMj0HtKvzZOr8wu6tF0YfFzonMdTWEziPcPn6Ham2ymVuue32VJ9l917vWmaIa53W++IPnUMZiq3yXMjVcJ5TuSbdFIp/557R552GzicNPS60zUL9QutZuU7BsRDY1velm6RxhXJJmpv9c2e8U7vvAtoLzj8Y7XPgkO+c9OMHHVevB1Dor96FwvsXxC9MDgBoK0W7zHrKd9H0N941EO1T2B2/SDsADM4cdvVr6Y9/nz3tQON/510qEf19xWg/2+PYftN9OT79rGPRPid6211jHX/N8dqD7/lq7fO9Dgbn+NZzyLFdO172fSw98dbBaJ/Cft/rxQqubhiaEV/P4tHqvXfUinlz/FA856w9/r4FAMVZ/dE+QwO+J+LIGzNd/ejJced7/dAM3/tgyeL5237Qt54Dc+LbtuWE8z3V2W3Y8X+59Xh1crzx/xuIiIiIiIhILmgCKiIiIiIiIpnQBFREREREREQyEZ2AklxC8mGSO0g+R/IzSfsXSe4muT35uaL24YqIiIiIiEiz8pztPQTgJjN7kmQXgJ+TfCi5b4OZ/efJBOApUhMqWrP8xm3ptm3xgkOT4S2yE4pt4bbRRTJCYy2E70T+kFCxJZdA4aOgq0ff9G7X4Hp2p9czWCyqosBQx6re9AI2uMIQEREREZEGEN0DamZ7zezJ5O9eAM8DWFTrwEQmi+RdJPeTfHZE2/9F8gWST5PcTHL2GI/dRfKZZO/+E9lFLeKnHJe8U45L3inHZSo6rXNASZ4D4N0AHk2aPp28OO4iOafKsYlM1iYAl1e0PQTgV8zsnQD+CcAfj/P495nZCjNbWaP4RCZrE5Tjkm+boByXfNsE5bhMMe4JKMkZAL4P4LNmdhTA1wAsA7ACwF4AXxnjcetIPkHyib4jfVUIWcTHzLoBHK5o+7GZnbow3zYAizMPTKRKlOOSd8pxyTvluExFrgkoySLKk89vmtl9AGBm+8xs2MxKAL4O4KLQY83sDjNbaWYrp832XfhaJCP/AcAPx7jPAPyY5M9JrsswJpFqUo5L3inHJe+U45I70SJEJAngTgDPm9lXR7SfbWZ7k5tXAXg29PiYiRYJOnHVxYFWXyGenVdvHHV79Q3XpfqECgLtWcNUW7DIzlXpx3ZvTT/WM36Ip3BTqE9Ix+ZH450ArF2/etTtUCGh0Fg9G1a5xvcIjRWKY5djLJJ/gnKBrW+O0eW9Zrab5BkAHiL5QvItZWisdQDWAUDnWZ2OpQN3L93i6pc1b1xrX14d79TAy5wKqpXjI/O7MHc2Bg7Fv0hsaY8XVFtw/sFoHwA4cKgr2qf40nTXWIMzS65+HsWjvvfrwZm+8dp3t0X7dL067BpruD0+VttRX9G7wxc5t1l//Pvsoc6JF9oLqUWOt86eA/4ynk90rErbOcfinQD0nXA8X79sd4015HspoNTmWAHv0+XZGABaDhWjfdoP+8YqFeNjtR11DYXec32vK1g8xwdmNUeOt7zq2CHkeIuz+YPxTgCG+uM1T1tfi78OAKBUcHWDtTqeC++/BOfxoi0n4sG1HfH97+BQfKwW3+ZH/zzfipaK8diGZlQnxz2b9DcBfALA+ysuufKXyYnPTwN4H4AbqxKRSI2R/D0AvwXgd80s+Eoys93J7/0ANmOMPfxJH+3ll4ZSzRwfmd+FGb4vWERqrWY53qkcl8ZQqxxvUY5LA4h+JWFmjyD8PcgD1Q9HpLZIXg7gjwD8KzM7MUafTgAtZtab/P1hAF/KMEyRCVOOS94pxyXvlOOSd6dVBVekmZD8NoCtAM4n+SrJTwG4FUAXyoeqbCe5Mem7kOSpL1XOBPAIyacAPAbgf5jZj+qwCiLjUo5L3inHJe+U4zIVxQ/KFmlSZnZNoPnOMfruAXBF8vdLAN5Vw9BEqkI5LnmnHJe8U47LVJTpBLT3+PRUsZzlN25L9Vu4bXTRiT2relN9QgVptgSKmay+5YJU29pLRhczCRX/CRW3CbWFHrt8fboY0vKK2z2BuEKPC/XrRqDwUUVsnj4AsGeDb90riyiFijSFC0OlBbeZ43GeXAGAf7jPFYaIiIiIiGRMh+CKiIiIiIhIJjQBFRERERERkUxoAioiIiIiIiKZ0ARUREREREREMpFpEaKuzpNYc8lzo9q6A8WEerZWNGxIj7Xz6o2ptrUvr061bbnt9lTbsnuvHz9QhAvlhJa5+obrUm2hAkCVQkV8QgWHvHEsQ3ydOjY/GojDVziocpmru9PrHSqiBGexpZDKYkihwlML8VyqTfLp7kCRsUqh9wCpnpYBouPV+L+NE4uH4n0Gir6FHmyPdulfEF/e6SjFF4niUd+/z7Z5fa5+g12FaJ/D/dN9Y80MXrO+Ynmhy3unte9uc/Ur9Mf7nFha3eepJkpA67H4tumfX4r2GXxjmmuRrQfir4WS72mAFeLPPQDA0Y8lX460zXY8+QBsVrxP72Cna6zh6fHtbwVf/NNei7/2AKBlON7nxCJHpzqjAa19jhyf49jGfb5tVzgSf790Zq4/xz18KYKWGYOuftYRH/DkgOMfDABzxNZ60jUU2l6v3v7G/nnxvPDQHlARERERERHJhCagIiIiIiIikglNQEVERERERCQTmoCKiIiIiIhIJjItQtT32vRUoZ1QMZ5KoUI8oUJCy2/clmpbfVW6WE7lMkNFcUIFgUKFd0KxVRZaCvEUKhpLqNBK5TJDBVvWXpJ+XPdW3xnYlcsMrXeo4NBk+nlyo3vrxLejiIiIiIhkS3tARUREREREJBOagIqIiIiIiEgmohNQktNIPkbyKZLPkfzTpP2tJB8l2UPyXpLOq1SJiIiIiIjIVOQ5B7QfwPvN7BjJIoBHSP4QwB8C2GBm3yG5EcCnAHxtvIFajhxHx+ZHR7Ut3NaV6ld5Xt/C7vS5gKHzNu/esz3VtuzeVeOFVBY4JzFky223p9pW35A+L7SnO36OY+g80T2retNjbUjH7znvcVmgj+fcVCC8vSvPWQ318Z4Pu+fG9HouxMWu2BrdOW3HguffVgqdx1vJMw4QPh+60s6rN7rGqgfPtsjDMvOgNL2EvnfEr3ztObSmo813Ye/+E/Hz1Aemu4ZCy0nfQT/TDsT79S1wXox7wHeh9tLxYtWW2erYZtbqu5j7jF+6umGo01FPoK06FzCvJZtm6Du/L9qPLfHtN619yLXMwcPx5364w7ftCsd9OV485MjxM5z55kslDA/FXwul2cOusTzr2TLoq3HRsd+3AoMdjvGa4LhCKxr6FsXff9kefy4KBeeT75huDE/3jVXo9z2vhZPxfoOzfDne4ni9A4A5Qhvqcub4iXgyccj5P+2wL/6h6fEV6K9SjkeHsbJjyc1i8mMA3g/ge0n7PQB+pzohiVQPybtI7if57Ii2L5LcTXJ78nPFGI+9nOQvkr38N2cXtYiP8lvyTjkueaccl6nINY8lWSC5HcB+AA8B2AngiJmd+nrvVQCLahOiyKRsAnB5oH2Dma1Ifh6ovJNkAcBtAD4C4AIA15D07SoXyc4mKL8l3zZBOS75tgnKcZliXBNQMxs2sxUAFgO4CMDbvQsguY7kEySfGET/BMMUmRgz6wZweAIPvQhAj5m9ZGYDAL4D4MqqBicyScpvyTvluOSdclymotM6ktfMjgB4GMAlAGaTPHVQ92IAu8d4zB1mttLMVhbRPqlgRaro0ySfTg59mRO4fxGAV0bc1l5+aSbKb8k75bjknXJccit6VjDJBQAGzewIyekAPgTgFpQnoh9F+RuXawH8IDbW2955Ag8+OLpQUKiASqpwzSXpsUKFeNbGAhiDt3hOsNjLGt8yKov2hAoVLd+WLqy0EL7CQZXbIxR/T6DY0pr16X6VBYdCQkWgQsWcggWZNqS3d6ioUaXQOnkKMgV8DcCfoXwu858B+AqA/zCRgYDyXn4A6wBg6SJPXS+RmqpZfhfmzapGfCKTVcMcn12N+EQmq3Y5Plc5LvXn2QN6NoCHST4N4HEAD5nZ3wNYD+APSfYAmAfgztqFKVI9ZrYvOay8BODrKB/GUmk3gCUjbrv28i+Y56t4KVIrtczvwszO6gcscpqU45J3Nc3xLuW41F90d42ZPQ3g3YH2lxB+QYg0NJJnm9ne5OZVAJ4NdHscwHkk34ryG/rHAHw8oxBFJkz5LXmnHJe8U45L3ul4Qck1kt8GcCmA+SRfBfAFAJeSXIHyoS27AFyX9F0I4K/N7AozGyL5aQAPAigAuMvMfMdDi2RE+S15pxyXvFOOy1SkCajkmpldE2gOHi5uZnsAXDHi9gMAUqXPRRqF8lvyTjkueaccl6mIZvHiL1VbGHkAwC8BzAdwMLMFV1+zxw80/zqMF/9bzGxBlsGcMiLHT2n27Qw0/zrkMf665Hggv4F8bt9m0uzxA+l1aKT3cKD5t7Hir6+GeQ8HlOMNKo/xj5vjmU5A31wo+YSZrcx8wVXS7PEDzb8OzRJ/s8Q5nmZfB8VfW40eX4zir79GX4dGjy9G8ddXM8TfDDGOR/HX10TiP63rgIqIiIiIiIhMlCagIiIiIiIikol6TUDvqNNyq6XZ4weafx2aJf5miXM8zb4Oir+2Gj2+GMVff42+Do0eX4zir69miL8ZYhyP4q+v046/LueAioiIiIiIyNSjQ3BFREREREQkE5lPQEleTvIXJHtI3pz18k8XybtI7if57Ii2uSQfIvli8ntOPWMcD8klJB8muYPkcyQ/k7Q3xTqQnEbyMZJPJfH/adL+VpKPJnl0L8m2esdaqdlyvRLJXSSfIbmd5BP1jiem2V+rwJjr8EWSu5PnYTvJK8YbI0vK8Ww1e44rv7OnHM+WcjxbzZbfgHL8lEwnoCQLAG4D8BEAFwC4huQFWcYwAZsAXF7RdjOAn5rZeQB+mtxuVEMAbjKzCwCsAnBDss2bZR36AbzfzN4FYAWAy0muAnALgA1mthzA6wA+VccYU5o010PeZ2YrmqQ8+CY092sVCK8DUM71FclPQ1x0XDleF5vQ3Dm+CcrvelCOZ2cTlONZa6b8BpTjALLfA3oRgB4ze8nMBgB8B8CVGcdwWsysG8DhiuYrAdyT/H0PgN/JNKjTYGZ7zezJ5O9eAM8DWIQmWQcrO5bcLCY/BuD9AL6XtDdi/E2X682u2V+rwJjr0KiU4xlr9hxXfkuMcjxTyvE6UI6XZT0BXQTglRG3X03ams2ZZrY3+fs1AGfWMxgvkucAeDeAR9FE60CyQHI7gP0AHgKwE8ARMxtKujRiHuUh1w3Aj0n+nOS6egczQU2T5xGfJvl0cuhLoxyaoxxvDHnIceV37SjHG4NyvDbykN/AFMxxFSGaJCuXEW74UsIkZwD4PoDPmtnRkfc1+jqY2bCZrQCwGOVv7N5e55Cmivea2XtQPjznBpJr6h3QZDR6no/jawCWoXwI+l4AX6lvOLmiHK8/5XdtKcfrTzleO7nKb2Dq5HjWE9DdAJaMuL04aWs2+0ieDQDJ7/11jmdcJIsoTz6/aWb3Jc1NtQ4AYGZHADwM4BIAs0m2Jnc1Yh41fa6b2e7k934Am1Ge/DebpsvzSma2L/kSpgTg62ic50E53hiaOseV37WlHK8/5Xjt5CS/gSmY41lPQB8HcB7LFUzbAHwMwP0Zx1AN9wO4Nvn7WgA/qGMs4yJJAHcCeN7MvjrirqZYB5ILSM5O/p4O4EMon8f6MICPJt0aMf6mznWSnSS7Tv0N4MMAnh3/UQ2pKfJ8PKf+KSWuQuM8D8rxxtDUOa78rh3leGNQjtdGjvIbmIo5bmaZ/gC4AsA/oXwe359kvfwJxPttlHcnD6J8fPynAMxDuUrViwB+AmBuveMcJ/73orwr/2kA25OfK5plHQC8E8A/JvE/C+DzSfu5AB4D0APgbwG01zvWQOxNlesVsZ8L4Knk57lmiL/ZX6vjrMPfAHgmeQ3cD+Dsesc5Il7leP3zo2lyXPmdefzK8caIXzlem9ibLr/HyZEpl+NMBhMRERERERGpKRUhEhERERERkUxoAioiIiIiIiKZ0ARUREREREREMqEJqIiIiIiIiGRCE1ARERERERHJhCagIiIiIiIikglNQEVERERERCQTmoCKiIiIiIhIJv5/lHa/9IwBczEAAAAASUVORK5CYII=\n",
      "text/plain": [
       "<Figure size 1152x216 with 5 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "from layers import Blur2D\n",
    "\n",
    "x = x_in = Input(shape=(32, 32, 10))\n",
    "x1 = MaxPool2D(3, 2, padding='same')(x)\n",
    "x2 = AvgPool2D(3, 2, padding='same')(x)\n",
    "x3 = Blur2D(5, 2, padding='same')(x)\n",
    "x4 = Blur2D(7, 2, padding='same')(x)\n",
    "\n",
    "model = Model(x_in, [x1,x2,x3,x4])\n",
    "\n",
    "batch = np.random.binomial(4, 0.5, size=(4,32,32,10)) / 4\n",
    "batch[:,:16,16:] = 0\n",
    "batch[:,4:8,20:24] = 1\n",
    "\n",
    "preds = model.predict(batch)\n",
    "\n",
    "plt.figure(figsize=(16,3))\n",
    "plt.subplot(151); plt.title('input')\n",
    "plt.imshow(batch[0,:,:,0], vmin=0, vmax=1)\n",
    "plt.subplot(152); plt.title('max 3')\n",
    "plt.imshow(preds[0][0,:,:,0], vmin=0, vmax=1)\n",
    "plt.subplot(153); plt.title('avg 3')\n",
    "plt.imshow(preds[1][0,:,:,0], vmin=0, vmax=1)\n",
    "plt.subplot(154); plt.title('blur 5')\n",
    "plt.imshow(preds[2][0,:,:,0], vmin=0, vmax=1)\n",
    "plt.subplot(155); plt.title('blur 7')\n",
    "plt.imshow(preds[3][0,:,:,0], vmin=0, vmax=1)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.8.10"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {},
   "toc_section_display": true,
   "toc_window_display": false
  },
  "varInspector": {
   "cols": {
    "lenName": 16,
    "lenType": 16,
    "lenVar": 40
   },
   "kernels_config": {
    "python": {
     "delete_cmd_postfix": "",
     "delete_cmd_prefix": "del ",
     "library": "var_list.py",
     "varRefreshCmd": "print(var_dic_list())"
    },
    "r": {
     "delete_cmd_postfix": ") ",
     "delete_cmd_prefix": "rm(",
     "library": "var_list.r",
     "varRefreshCmd": "cat(var_dic_list()) "
    }
   },
   "types_to_exclude": [
    "module",
    "function",
    "builtin_function_or_method",
    "instance",
    "_Feature"
   ],
   "window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}