train_shapes.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "# Mask R-CNN - Train on Shapes Dataset\n",
    "\n",
    "\n",
    "This notebook shows how to train Mask R-CNN on your own dataset. To keep things simple we use a synthetic dataset of shapes (squares, triangles, and circles) which enables fast training. You'd still need a GPU, though, because the network backbone is a Resnet101, which would be too slow to train on a CPU. On a GPU, you can start to get okay-ish results in a few minutes, and good results in less than an hour.\n",
    "\n",
    "The code of the *Shapes* dataset is included below. It generates images on the fly, so it doesn't require downloading any data. And it can generate images of any size, so we pick a small image size to train faster. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/orestis/anaconda3/lib/python3.6/site-packages/matplotlib/font_manager.py:280: UserWarning: Matplotlib is building the font cache using fc-list. This may take a moment.\n",
      "  'Matplotlib is building the font cache using fc-list. '\n",
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import os\n",
    "import sys\n",
    "import random\n",
    "import math\n",
    "import re\n",
    "import time\n",
    "import numpy as np\n",
    "import cv2\n",
    "import matplotlib\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "from config import Config\n",
    "import utils\n",
    "import model as modellib\n",
    "import visualize\n",
    "from model import log\n",
    "\n",
    "%matplotlib inline \n",
    "\n",
    "# Root directory of the project\n",
    "ROOT_DIR = os.getcwd()\n",
    "\n",
    "# Directory to save logs and trained model\n",
    "MODEL_DIR = os.path.join(ROOT_DIR, \"logs\")\n",
    "\n",
    "# Local path to trained weights file\n",
    "COCO_MODEL_PATH = os.path.join(ROOT_DIR, \"mask_rcnn_coco.h5\")\n",
    "# Download COCO trained weights from Releases if needed\n",
    "if not os.path.exists(COCO_MODEL_PATH):\n",
    "    utils.download_trained_weights(COCO_MODEL_PATH)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Configurations"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "Configurations:\n",
      "BACKBONE_SHAPES                [[32 32]\n",
      " [16 16]\n",
      " [ 8  8]\n",
      " [ 4  4]\n",
      " [ 2  2]]\n",
      "BACKBONE_STRIDES               [4, 8, 16, 32, 64]\n",
      "BATCH_SIZE                     8\n",
      "BBOX_STD_DEV                   [ 0.1  0.1  0.2  0.2]\n",
      "DETECTION_MAX_INSTANCES        100\n",
      "DETECTION_MIN_CONFIDENCE       0.7\n",
      "DETECTION_NMS_THRESHOLD        0.3\n",
      "GPU_COUNT                      1\n",
      "IMAGES_PER_GPU                 8\n",
      "IMAGE_MAX_DIM                  128\n",
      "IMAGE_MIN_DIM                  128\n",
      "IMAGE_PADDING                  True\n",
      "IMAGE_SHAPE                    [128 128   3]\n",
      "LEARNING_MOMENTUM              0.9\n",
      "LEARNING_RATE                  0.002\n",
      "MASK_POOL_SIZE                 14\n",
      "MASK_SHAPE                     [28, 28]\n",
      "MAX_GT_INSTANCES               100\n",
      "MEAN_PIXEL                     [ 123.7  116.8  103.9]\n",
      "MINI_MASK_SHAPE                (56, 56)\n",
      "NAME                           shapes\n",
      "NUM_CLASSES                    4\n",
      "POOL_SIZE                      7\n",
      "POST_NMS_ROIS_INFERENCE        1000\n",
      "POST_NMS_ROIS_TRAINING         2000\n",
      "ROI_POSITIVE_RATIO             0.33\n",
      "RPN_ANCHOR_RATIOS              [0.5, 1, 2]\n",
      "RPN_ANCHOR_SCALES              (8, 16, 32, 64, 128)\n",
      "RPN_ANCHOR_STRIDE              2\n",
      "RPN_BBOX_STD_DEV               [ 0.1  0.1  0.2  0.2]\n",
      "RPN_TRAIN_ANCHORS_PER_IMAGE    256\n",
      "STEPS_PER_EPOCH                100\n",
      "TRAIN_ROIS_PER_IMAGE           32\n",
      "USE_MINI_MASK                  True\n",
      "USE_RPN_ROIS                   True\n",
      "VALIDATION_STEPS               50\n",
      "WEIGHT_DECAY                   0.0001\n",
      "\n",
      "\n"
     ]
    }
   ],
   "source": [
    "class ShapesConfig(Config):\n",
    "    \"\"\"Configuration for training on the toy shapes dataset.\n",
    "    Derives from the base Config class and overrides values specific\n",
    "    to the toy shapes dataset.\n",
    "    \"\"\"\n",
    "    # Give the configuration a recognizable name\n",
    "    NAME = \"shapes\"\n",
    "\n",
    "    # Train on 1 GPU and 8 images per GPU. We can put multiple images on each\n",
    "    # GPU because the images are small. Batch size is 8 (GPUs * images/GPU).\n",
    "    GPU_COUNT = 1\n",
    "    IMAGES_PER_GPU = 8\n",
    "\n",
    "    # Number of classes (including background)\n",
    "    NUM_CLASSES = 1 + 3  # background + 3 shapes\n",
    "\n",
    "    # Use small images for faster training. Set the limits of the small side\n",
    "    # the large side, and that determines the image shape.\n",
    "    IMAGE_MIN_DIM = 128\n",
    "    IMAGE_MAX_DIM = 128\n",
    "\n",
    "    # Use smaller anchors because our image and objects are small\n",
    "    RPN_ANCHOR_SCALES = (8, 16, 32, 64, 128)  # anchor side in pixels\n",
    "\n",
    "    # Reduce training ROIs per image because the images are small and have\n",
    "    # few objects. Aim to allow ROI sampling to pick 33% positive ROIs.\n",
    "    TRAIN_ROIS_PER_IMAGE = 32\n",
    "\n",
    "    # Use a small epoch since the data is simple\n",
    "    STEPS_PER_EPOCH = 100\n",
    "\n",
    "    # use small validation steps since the epoch is small\n",
    "    VALIDATION_STEPS = 5\n",
    "    \n",
    "config = ShapesConfig()\n",
    "config.display()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Notebook Preferences"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "def get_ax(rows=1, cols=1, size=8):\n",
    "    \"\"\"Return a Matplotlib Axes array to be used in\n",
    "    all visualizations in the notebook. Provide a\n",
    "    central point to control graph sizes.\n",
    "    \n",
    "    Change the default size attribute to control the size\n",
    "    of rendered images\n",
    "    \"\"\"\n",
    "    _, ax = plt.subplots(rows, cols, figsize=(size*cols, size*rows))\n",
    "    return ax"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Dataset\n",
    "\n",
    "Create a synthetic dataset\n",
    "\n",
    "Extend the Dataset class and add a method to load the shapes dataset, `load_shapes()`, and override the following methods:\n",
    "\n",
    "* load_image()\n",
    "* load_mask()\n",
    "* image_reference()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "class ShapesDataset(utils.Dataset):\n",
    "    \"\"\"Generates the shapes synthetic dataset. The dataset consists of simple\n",
    "    shapes (triangles, squares, circles) placed randomly on a blank surface.\n",
    "    The images are generated on the fly. No file access required.\n",
    "    \"\"\"\n",
    "\n",
    "    def load_shapes(self, count, height, width):\n",
    "        \"\"\"Generate the requested number of synthetic images.\n",
    "        count: number of images to generate.\n",
    "        height, width: the size of the generated images.\n",
    "        \"\"\"\n",
    "        # Add classes\n",
    "        self.add_class(\"shapes\", 1, \"square\")\n",
    "        self.add_class(\"shapes\", 2, \"circle\")\n",
    "        self.add_class(\"shapes\", 3, \"triangle\")\n",
    "\n",
    "        # Add images\n",
    "        # Generate random specifications of images (i.e. color and\n",
    "        # list of shapes sizes and locations). This is more compact than\n",
    "        # actual images. Images are generated on the fly in load_image().\n",
    "        for i in range(count):\n",
    "            bg_color, shapes = self.random_image(height, width)\n",
    "            self.add_image(\"shapes\", image_id=i, path=None,\n",
    "                           width=width, height=height,\n",
    "                           bg_color=bg_color, shapes=shapes)\n",
    "\n",
    "    def load_image(self, image_id):\n",
    "        \"\"\"Generate an image from the specs of the given image ID.\n",
    "        Typically this function loads the image from a file, but\n",
    "        in this case it generates the image on the fly from the\n",
    "        specs in image_info.\n",
    "        \"\"\"\n",
    "        info = self.image_info[image_id]\n",
    "        bg_color = np.array(info['bg_color']).reshape([1, 1, 3])\n",
    "        image = np.ones([info['height'], info['width'], 3], dtype=np.uint8)\n",
    "        image = image * bg_color.astype(np.uint8)\n",
    "        for shape, color, dims in info['shapes']:\n",
    "            image = self.draw_shape(image, shape, dims, color)\n",
    "        return image\n",
    "\n",
    "    def image_reference(self, image_id):\n",
    "        \"\"\"Return the shapes data of the image.\"\"\"\n",
    "        info = self.image_info[image_id]\n",
    "        if info[\"source\"] == \"shapes\":\n",
    "            return info[\"shapes\"]\n",
    "        else:\n",
    "            super(self.__class__).image_reference(self, image_id)\n",
    "\n",
    "    def load_mask(self, image_id):\n",
    "        \"\"\"Generate instance masks for shapes of the given image ID.\n",
    "        \"\"\"\n",
    "        info = self.image_info[image_id]\n",
    "        shapes = info['shapes']\n",
    "        count = len(shapes)\n",
    "        mask = np.zeros([info['height'], info['width'], count], dtype=np.uint8)\n",
    "        for i, (shape, _, dims) in enumerate(info['shapes']):\n",
    "            mask[:, :, i:i+1] = self.draw_shape(mask[:, :, i:i+1].copy(),\n",
    "                                                shape, dims, 1)\n",
    "        # Handle occlusions\n",
    "        occlusion = np.logical_not(mask[:, :, -1]).astype(np.uint8)\n",
    "        for i in range(count-2, -1, -1):\n",
    "            mask[:, :, i] = mask[:, :, i] * occlusion\n",
    "            occlusion = np.logical_and(occlusion, np.logical_not(mask[:, :, i]))\n",
    "        # Map class names to class IDs.\n",
    "        class_ids = np.array([self.class_names.index(s[0]) for s in shapes])\n",
    "        return mask, class_ids.astype(np.int32)\n",
    "\n",
    "    def draw_shape(self, image, shape, dims, color):\n",
    "        \"\"\"Draws a shape from the given specs.\"\"\"\n",
    "        # Get the center x, y and the size s\n",
    "        x, y, s = dims\n",
    "        if shape == 'square':\n",
    "            cv2.rectangle(image, (x-s, y-s), (x+s, y+s), color, -1)\n",
    "        elif shape == \"circle\":\n",
    "            cv2.circle(image, (x, y), s, color, -1)\n",
    "        elif shape == \"triangle\":\n",
    "            points = np.array([[(x, y-s),\n",
    "                                (x-s/math.sin(math.radians(60)), y+s),\n",
    "                                (x+s/math.sin(math.radians(60)), y+s),\n",
    "                                ]], dtype=np.int32)\n",
    "            cv2.fillPoly(image, points, color)\n",
    "        return image\n",
    "\n",
    "    def random_shape(self, height, width):\n",
    "        \"\"\"Generates specifications of a random shape that lies within\n",
    "        the given height and width boundaries.\n",
    "        Returns a tuple of three valus:\n",
    "        * The shape name (square, circle, ...)\n",
    "        * Shape color: a tuple of 3 values, RGB.\n",
    "        * Shape dimensions: A tuple of values that define the shape size\n",
    "                            and location. Differs per shape type.\n",
    "        \"\"\"\n",
    "        # Shape\n",
    "        shape = random.choice([\"square\", \"circle\", \"triangle\"])\n",
    "        # Color\n",
    "        color = tuple([random.randint(0, 255) for _ in range(3)])\n",
    "        # Center x, y\n",
    "        buffer = 20\n",
    "        y = random.randint(buffer, height - buffer - 1)\n",
    "        x = random.randint(buffer, width - buffer - 1)\n",
    "        # Size\n",
    "        s = random.randint(buffer, height//4)\n",
    "        return shape, color, (x, y, s)\n",
    "\n",
    "    def random_image(self, height, width):\n",
    "        \"\"\"Creates random specifications of an image with multiple shapes.\n",
    "        Returns the background color of the image and a list of shape\n",
    "        specifications that can be used to draw the image.\n",
    "        \"\"\"\n",
    "        # Pick random background color\n",
    "        bg_color = np.array([random.randint(0, 255) for _ in range(3)])\n",
    "        # Generate a few random shapes and record their\n",
    "        # bounding boxes\n",
    "        shapes = []\n",
    "        boxes = []\n",
    "        N = random.randint(1, 4)\n",
    "        for _ in range(N):\n",
    "            shape, color, dims = self.random_shape(height, width)\n",
    "            shapes.append((shape, color, dims))\n",
    "            x, y, s = dims\n",
    "            boxes.append([y-s, x-s, y+s, x+s])\n",
    "        # Apply non-max suppression wit 0.3 threshold to avoid\n",
    "        # shapes covering each other\n",
    "        keep_ixs = utils.non_max_suppression(np.array(boxes), np.arange(N), 0.3)\n",
    "        shapes = [s for i, s in enumerate(shapes) if i in keep_ixs]\n",
    "        return bg_color, shapes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Training dataset\n",
    "dataset_train = ShapesDataset()\n",
    "dataset_train.load_shapes(500, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1])\n",
    "dataset_train.prepare()\n",
    "\n",
    "# Validation dataset\n",
    "dataset_val = ShapesDataset()\n",
    "dataset_val.load_shapes(50, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1])\n",
    "dataset_val.prepare()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAy4AAACnCAYAAAD35AgmAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f3e028cb278>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAy4AAACnCAYAAAD35AgmAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAABAe8IFAABoT7gAAADt/S/gbLui+LpWmgAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f3e00e5ecc0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAy4AAACnCAYAAAD35AgmAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nXAAAgO79H5CYub/IHqsBAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f3e00c32a20>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAy4AAACnCAYAAAD35AgmAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f3e009f1dd8>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Load and display random samples\n",
    "image_ids = np.random.choice(dataset_train.image_ids, 4)\n",
    "for image_id in image_ids:\n",
    "    image = dataset_train.load_image(image_id)\n",
    "    mask, class_ids = dataset_train.load_mask(image_id)\n",
    "    visualize.display_top_masks(image, mask, class_ids, dataset_train.class_names)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Ceate Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Create model in training mode\n",
    "model = modellib.MaskRCNN(mode=\"training\", config=config,\n",
    "                          model_dir=MODEL_DIR)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true,
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "# Which weights to start with?\n",
    "init_with = \"coco\"  # imagenet, coco, or last\n",
    "\n",
    "if init_with == \"imagenet\":\n",
    "    model.load_weights(model.get_imagenet_weights(), by_name=True)\n",
    "elif init_with == \"coco\":\n",
    "    # Load weights trained on MS COCO, but skip layers that\n",
    "    # are different due to the different number of classes\n",
    "    # See README for instructions to download the COCO weights\n",
    "    model.load_weights(COCO_MODEL_PATH, by_name=True,\n",
    "                       exclude=[\"mrcnn_class_logits\", \"mrcnn_bbox_fc\", \n",
    "                                \"mrcnn_bbox\", \"mrcnn_mask\"])\n",
    "elif init_with == \"last\":\n",
    "    # Load the last model you trained and continue training\n",
    "    model.load_weights(model.find_last()[1], by_name=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Training\n",
    "\n",
    "Train in two stages:\n",
    "1. Only the heads. Here we're freezing all the backbone layers and training only the randomly initialized layers (i.e. the ones that we didn't use pre-trained weights from MS COCO). To train only the head layers, pass `layers='heads'` to the `train()` function.\n",
    "\n",
    "2. Fine-tune all layers. For this simple example it's not necessary, but we're including it to show the process. Simply pass `layers=\"all` to train all layers."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false,
    "deletable": true,
    "editable": true,
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "Starting at epoch 0. LR=0.002\n",
      "\n",
      "Checkpoint Path: /home/orestis/repositories/Mask_RCNN/logs/shapes20171105T2327/mask_rcnn_shapes_{epoch:04d}.h5\n",
      "Selecting layers to train\n",
      "fpn_c5p5               (Conv2D)\n",
      "fpn_c4p4               (Conv2D)\n",
      "fpn_c3p3               (Conv2D)\n",
      "fpn_c2p2               (Conv2D)\n",
      "fpn_p5                 (Conv2D)\n",
      "fpn_p2                 (Conv2D)\n",
      "fpn_p3                 (Conv2D)\n",
      "fpn_p4                 (Conv2D)\n",
      "In model:  rpn_model\n",
      "    rpn_conv_shared        (Conv2D)\n",
      "    rpn_class_raw          (Conv2D)\n",
      "    rpn_bbox_pred          (Conv2D)\n",
      "mrcnn_mask_conv1       (TimeDistributed)\n",
      "mrcnn_mask_bn1         (TimeDistributed)\n",
      "mrcnn_mask_conv2       (TimeDistributed)\n",
      "mrcnn_class_conv1      (TimeDistributed)\n",
      "mrcnn_mask_bn2         (TimeDistributed)\n",
      "mrcnn_class_bn1        (TimeDistributed)\n",
      "mrcnn_mask_conv3       (TimeDistributed)\n",
      "mrcnn_mask_bn3         (TimeDistributed)\n",
      "mrcnn_class_conv2      (TimeDistributed)\n",
      "mrcnn_class_bn2        (TimeDistributed)\n",
      "mrcnn_mask_conv4       (TimeDistributed)\n",
      "mrcnn_mask_bn4         (TimeDistributed)\n",
      "mrcnn_bbox_fc          (TimeDistributed)\n",
      "mrcnn_mask_deconv      (TimeDistributed)\n",
      "mrcnn_class_logits     (TimeDistributed)\n",
      "mrcnn_mask             (TimeDistributed)\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/orestis/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/gradients_impl.py:95: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.\n",
      "  \"Converting sparse IndexedSlices to a dense Tensor of unknown shape. \"\n",
      "/home/orestis/anaconda3/lib/python3.6/site-packages/keras/engine/training.py:1987: UserWarning: Using a generator with `use_multiprocessing=True` and multiple workers may duplicate your data. Please consider using the`keras.utils.Sequence class.\n",
      "  UserWarning('Using a generator with `use_multiprocessing=True`'\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/1\n",
      " 74/100 [=====================>........] - ETA: 1072s - loss: 2.9590 - rpn_class_loss: 0.0274 - rpn_bbox_loss: 1.1385 - mrcnn_class_loss: 0.4422 - mrcnn_bbox_loss: 0.4137 - mrcnn_mask_loss: 0.5959"
     ]
    }
   ],
   "source": [
    "# Train the head branches\n",
    "# Passing layers=\"heads\" freezes all layers except the head\n",
    "# layers. You can also pass a regular expression to select\n",
    "# which layers to train by name pattern.\n",
    "model.train(dataset_train, dataset_val, \n",
    "            learning_rate=config.LEARNING_RATE, \n",
    "            epochs=1, \n",
    "            layers='heads')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true,
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "# Fine tune all layers\n",
    "# Passing layers=\"all\" trains all layers. You can also \n",
    "# pass a regular expression to select which layers to\n",
    "# train by name pattern.\n",
    "model.train(dataset_train, dataset_val, \n",
    "            learning_rate=config.LEARNING_RATE / 10,\n",
    "            epochs=2, \n",
    "            layers=\"all\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Save weights\n",
    "# Typically not needed because callbacks save after every epoch\n",
    "# Uncomment to save manually\n",
    "# model_path = os.path.join(MODEL_DIR, \"mask_rcnn_shapes.h5\")\n",
    "# model.keras_model.save_weights(model_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Detection"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "class InferenceConfig(ShapesConfig):\n",
    "    GPU_COUNT = 1\n",
    "    IMAGES_PER_GPU = 1\n",
    "\n",
    "inference_config = InferenceConfig()\n",
    "\n",
    "# Recreate the model in inference mode\n",
    "model = modellib.MaskRCNN(mode=\"inference\", \n",
    "                          config=inference_config,\n",
    "                          model_dir=MODEL_DIR)\n",
    "\n",
    "# Get path to saved weights\n",
    "# Either set a specific path or find last trained weights\n",
    "# model_path = os.path.join(ROOT_DIR, \".h5 file name here\")\n",
    "model_path = model.find_last()[1]\n",
    "\n",
    "# Load trained weights (fill in path to trained weights here)\n",
    "assert model_path != \"\", \"Provide path to trained weights\"\n",
    "print(\"Loading weights from \", model_path)\n",
    "model.load_weights(model_path, by_name=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Test on a random image\n",
    "image_id = random.choice(dataset_val.image_ids)\n",
    "original_image, image_meta, gt_class_id, gt_bbox, gt_mask =\\\n",
    "    modellib.load_image_gt(dataset_val, inference_config, \n",
    "                           image_id, use_mini_mask=False)\n",
    "\n",
    "log(\"original_image\", original_image)\n",
    "log(\"image_meta\", image_meta)\n",
    "log(\"gt_class_id\", gt_class_id)\n",
    "log(\"gt_bbox\", gt_bbox)\n",
    "log(\"gt_mask\", gt_mask)\n",
    "\n",
    "visualize.display_instances(original_image, gt_bbox, gt_mask, gt_class_id, \n",
    "                            dataset_train.class_names, figsize=(8, 8))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "results = model.detect([original_image], verbose=1)\n",
    "\n",
    "r = results[0]\n",
    "visualize.display_instances(original_image, r['rois'], r['masks'], r['class_ids'], \n",
    "                            dataset_val.class_names, r['scores'], ax=get_ax())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "deletable": true,
    "editable": true
   },
   "source": [
    "## Evaluation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": [
    "# Compute VOC-Style mAP @ IoU=0.5\n",
    "# Running on 10 images. Increase for better accuracy.\n",
    "image_ids = np.random.choice(dataset_val.image_ids, 10)\n",
    "APs = []\n",
    "for image_id in image_ids:\n",
    "    # Load image and ground truth data\n",
    "    image, image_meta, gt_class_id, gt_bbox, gt_mask =\\\n",
    "        modellib.load_image_gt(dataset_val, inference_config,\n",
    "                               image_id, use_mini_mask=False)\n",
    "    molded_images = np.expand_dims(modellib.mold_image(image, inference_config), 0)\n",
    "    # Run object detection\n",
    "    results = model.detect([image], verbose=0)\n",
    "    r = results[0]\n",
    "    # Compute AP\n",
    "    AP, precisions, recalls, overlaps =\\\n",
    "        utils.compute_ap(gt_bbox, gt_class_id, gt_mask,\n",
    "                         r[\"rois\"], r[\"class_ids\"], r[\"scores\"], r['masks'])\n",
    "    APs.append(AP)\n",
    "    \n",
    "print(\"mAP: \", np.mean(APs))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true,
    "deletable": true,
    "editable": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.6.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}