Unet-model.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Unet model\n",
    "example to get the Unet architecture without the cropping described in the original paper\n",
    "\n",
    "Reference code for architecture  \n",
    "https://github.com/zhixuhao/unet/blob/master/model.py  \n",
    "  \n",
    "  \n",
    "Reference model paper:  \n",
    "https://arxiv.org/abs/1505.04597  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/bsaldivar/anaconda3/envs/keras-gpu-35/lib/python3.5/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  from ._conv import register_converters as _register_converters\n",
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import numpy as np\n",
    "import tensorflow as tf\n",
    "import keras"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.layers import *\n",
    "from keras.models import *\n",
    "from keras.optimizers import Adam,Adadelta\n",
    "\n",
    "def get_name(prefix,layer_counter,scope=\"CNN\"):\n",
    "    #If layer_counter does not exist create it\n",
    "    counter = layer_counter.get(prefix,0)\n",
    "    output = scope+\"_\"+prefix+\"_\"+str(counter)\n",
    "    counter+=1\n",
    "    layer_counter[prefix]=counter\n",
    "    return output\n",
    "\n",
    "def get_unet(input_size = (256,256,3),output_channels = 1,lr = 0.01,decay = 1e-6,model_name=\"Unet\",layers_depth = [64,128,256,512,1024]):\n",
    "    lc = {}\n",
    "    inputs = Input(input_size,name=\"Input\")\n",
    "    x = inputs\n",
    "    scope=\"CNN\"\n",
    "    layers = []\n",
    "    layers_depth_up = layers_depth[:-1]\n",
    "    layers_depth_up.reverse()\n",
    "    for layer_depth in layers_depth:\n",
    "        for _ in range(2):\n",
    "            x = Conv2D(layer_depth,3,padding = 'same',kernel_initializer = 'he_normal',name=get_name(\"Conv\",lc,scope=scope))(x)\n",
    "            x = BatchNormalization(name=get_name(\"batch_norm\",lc,scope=scope))(x)\n",
    "            x = Activation('relu',name=get_name(\"relu\",lc,scope=scope))(x)\n",
    "            if _ == 1:\n",
    "                layers.append(x)\n",
    "        if layer_depth!=layers_depth[-1]:\n",
    "            x = MaxPooling2D(pool_size=(2, 2),name=get_name(\"MaxPool\",lc,scope=scope))(x)\n",
    "    for i,layer_depth_up in enumerate(layers_depth_up):\n",
    "        x = UpSampling2D(size = (2,2),name=get_name(\"Upsampling\",lc,scope=scope))(x)\n",
    "        x = Concatenate(axis = 3)([layers[-(i+2)],x])\n",
    "        for _ in range(2):\n",
    "            x = Conv2D(layer_depth_up,3,padding = 'same',kernel_initializer = 'he_normal',name=get_name(\"Conv\",lc,scope=scope))(x)\n",
    "            x = BatchNormalization(name=get_name(\"batch_norm\",lc,scope=scope))(x)\n",
    "            x = Activation('relu',name=get_name(\"relu\",lc,scope=scope))(x)\n",
    "    #Make output\n",
    "    x = Conv2D(output_channels,3,padding = 'same',kernel_initializer = 'he_normal',name=get_name(\"Conv\",lc,scope=scope))(x)\n",
    "    x = BatchNormalization(name=get_name(\"batch_norm\",lc,scope=scope))(x)\n",
    "    x = Activation('sigmoid',name=\"Output\")(x)\n",
    "    model = Model(inputs=inputs,outputs=x,name=model_name)\n",
    "    optimizer=Adadelta(lr=lr, rho=0.95, epsilon=None, decay=decay)\n",
    "    model.compile(optimizer=optimizer,loss = 'binary_crossentropy', metrics = ['accuracy'])\n",
    "    return model\n",
    "\n",
    "\n",
    "\n",
    "unet = get_unet(input_size = (256,256,3),output_channels = 1,lr = 0.01,decay = 1e-6,model_name=\"Unet\",layers_depth = [64,128,256,512,1024])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "__________________________________________________________________________________________________\n",
      "Layer (type)                    Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      "Input (InputLayer)              (None, 256, 256, 3)  0                                            \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_0 (Conv2D)             (None, 256, 256, 64) 1792        Input[0][0]                      \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_0 (BatchNormaliz (None, 256, 256, 64) 256         CNN_Conv_0[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_0 (Activation)         (None, 256, 256, 64) 0           CNN_batch_norm_0[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_1 (Conv2D)             (None, 256, 256, 64) 36928       CNN_relu_0[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_1 (BatchNormaliz (None, 256, 256, 64) 256         CNN_Conv_1[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_1 (Activation)         (None, 256, 256, 64) 0           CNN_batch_norm_1[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_MaxPool_0 (MaxPooling2D)    (None, 128, 128, 64) 0           CNN_relu_1[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_2 (Conv2D)             (None, 128, 128, 128 73856       CNN_MaxPool_0[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_2 (BatchNormaliz (None, 128, 128, 128 512         CNN_Conv_2[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_2 (Activation)         (None, 128, 128, 128 0           CNN_batch_norm_2[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_3 (Conv2D)             (None, 128, 128, 128 147584      CNN_relu_2[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_3 (BatchNormaliz (None, 128, 128, 128 512         CNN_Conv_3[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_3 (Activation)         (None, 128, 128, 128 0           CNN_batch_norm_3[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_MaxPool_1 (MaxPooling2D)    (None, 64, 64, 128)  0           CNN_relu_3[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_4 (Conv2D)             (None, 64, 64, 256)  295168      CNN_MaxPool_1[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_4 (BatchNormaliz (None, 64, 64, 256)  1024        CNN_Conv_4[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_4 (Activation)         (None, 64, 64, 256)  0           CNN_batch_norm_4[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_5 (Conv2D)             (None, 64, 64, 256)  590080      CNN_relu_4[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_5 (BatchNormaliz (None, 64, 64, 256)  1024        CNN_Conv_5[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_5 (Activation)         (None, 64, 64, 256)  0           CNN_batch_norm_5[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_MaxPool_2 (MaxPooling2D)    (None, 32, 32, 256)  0           CNN_relu_5[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_6 (Conv2D)             (None, 32, 32, 512)  1180160     CNN_MaxPool_2[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_6 (BatchNormaliz (None, 32, 32, 512)  2048        CNN_Conv_6[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_6 (Activation)         (None, 32, 32, 512)  0           CNN_batch_norm_6[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_7 (Conv2D)             (None, 32, 32, 512)  2359808     CNN_relu_6[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_7 (BatchNormaliz (None, 32, 32, 512)  2048        CNN_Conv_7[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_7 (Activation)         (None, 32, 32, 512)  0           CNN_batch_norm_7[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_MaxPool_3 (MaxPooling2D)    (None, 16, 16, 512)  0           CNN_relu_7[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_8 (Conv2D)             (None, 16, 16, 1024) 4719616     CNN_MaxPool_3[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_8 (BatchNormaliz (None, 16, 16, 1024) 4096        CNN_Conv_8[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_8 (Activation)         (None, 16, 16, 1024) 0           CNN_batch_norm_8[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_9 (Conv2D)             (None, 16, 16, 1024) 9438208     CNN_relu_8[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_9 (BatchNormaliz (None, 16, 16, 1024) 4096        CNN_Conv_9[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_9 (Activation)         (None, 16, 16, 1024) 0           CNN_batch_norm_9[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Upsampling_0 (UpSampling2D) (None, 32, 32, 1024) 0           CNN_relu_9[0][0]                 \n",
      "__________________________________________________________________________________________________\n",
      "concatenate_25 (Concatenate)    (None, 32, 32, 1536) 0           CNN_relu_7[0][0]                 \n",
      "                                                                 CNN_Upsampling_0[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_10 (Conv2D)            (None, 32, 32, 512)  7078400     concatenate_25[0][0]             \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_10 (BatchNormali (None, 32, 32, 512)  2048        CNN_Conv_10[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_10 (Activation)        (None, 32, 32, 512)  0           CNN_batch_norm_10[0][0]          \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_11 (Conv2D)            (None, 32, 32, 512)  2359808     CNN_relu_10[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_11 (BatchNormali (None, 32, 32, 512)  2048        CNN_Conv_11[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_11 (Activation)        (None, 32, 32, 512)  0           CNN_batch_norm_11[0][0]          \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Upsampling_1 (UpSampling2D) (None, 64, 64, 512)  0           CNN_relu_11[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "concatenate_26 (Concatenate)    (None, 64, 64, 768)  0           CNN_relu_5[0][0]                 \n",
      "                                                                 CNN_Upsampling_1[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_12 (Conv2D)            (None, 64, 64, 256)  1769728     concatenate_26[0][0]             \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_12 (BatchNormali (None, 64, 64, 256)  1024        CNN_Conv_12[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_12 (Activation)        (None, 64, 64, 256)  0           CNN_batch_norm_12[0][0]          \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_13 (Conv2D)            (None, 64, 64, 256)  590080      CNN_relu_12[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_13 (BatchNormali (None, 64, 64, 256)  1024        CNN_Conv_13[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_13 (Activation)        (None, 64, 64, 256)  0           CNN_batch_norm_13[0][0]          \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Upsampling_2 (UpSampling2D) (None, 128, 128, 256 0           CNN_relu_13[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "concatenate_27 (Concatenate)    (None, 128, 128, 384 0           CNN_relu_3[0][0]                 \n",
      "                                                                 CNN_Upsampling_2[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_14 (Conv2D)            (None, 128, 128, 128 442496      concatenate_27[0][0]             \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_14 (BatchNormali (None, 128, 128, 128 512         CNN_Conv_14[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_14 (Activation)        (None, 128, 128, 128 0           CNN_batch_norm_14[0][0]          \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_15 (Conv2D)            (None, 128, 128, 128 147584      CNN_relu_14[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_15 (BatchNormali (None, 128, 128, 128 512         CNN_Conv_15[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_15 (Activation)        (None, 128, 128, 128 0           CNN_batch_norm_15[0][0]          \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Upsampling_3 (UpSampling2D) (None, 256, 256, 128 0           CNN_relu_15[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "concatenate_28 (Concatenate)    (None, 256, 256, 192 0           CNN_relu_1[0][0]                 \n",
      "                                                                 CNN_Upsampling_3[0][0]           \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_16 (Conv2D)            (None, 256, 256, 64) 110656      concatenate_28[0][0]             \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_16 (BatchNormali (None, 256, 256, 64) 256         CNN_Conv_16[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_16 (Activation)        (None, 256, 256, 64) 0           CNN_batch_norm_16[0][0]          \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_17 (Conv2D)            (None, 256, 256, 64) 36928       CNN_relu_16[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_17 (BatchNormali (None, 256, 256, 64) 256         CNN_Conv_17[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_relu_17 (Activation)        (None, 256, 256, 64) 0           CNN_batch_norm_17[0][0]          \n",
      "__________________________________________________________________________________________________\n",
      "CNN_Conv_18 (Conv2D)            (None, 256, 256, 1)  577         CNN_relu_17[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "CNN_batch_norm_18 (BatchNormali (None, 256, 256, 1)  4           CNN_Conv_18[0][0]                \n",
      "__________________________________________________________________________________________________\n",
      "Output (Activation)             (None, 256, 256, 1)  0           CNN_batch_norm_18[0][0]          \n",
      "==================================================================================================\n",
      "Total params: 31,403,013\n",
      "Trainable params: 31,391,235\n",
      "Non-trainable params: 11,778\n",
      "__________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "unet.summary()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "keras-gpu-35",
   "language": "python",
   "name": "keras-gpu-35"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.5.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}