Object detection Faster R-CNN.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "\tfrom tensorflow.contrib import slim as contrib_slim\n",
    "\t\n",
    "\tfrom object_detection.meta_architectures import faster_rcnn_meta_arch\n",
    "\tfrom object_detection.utils import variables_helper\n",
    "\tfrom nets import inception_resnet_v2\n",
    "\t\n",
    "\tslim = contrib_slim\n",
    "\t\n",
    "\t\n",
    "\tclass FasterRCNNInceptionResnetV2FeatureExtractor(\n",
    "\t    faster_rcnn_meta_arch.FasterRCNNFeatureExtractor):\n",
    "\t  \"\"\"Faster R-CNN with Inception Resnet v2 feature extractor implementation.\"\"\"\n",
    "\t\n",
    "\t  def __init__(self,\n",
    "\t               is_training,\n",
    "\t               first_stage_features_stride,\n",
    "\t               batch_norm_trainable=False,\n",
    "\t               reuse_weights=None,\n",
    "\t               weight_decay=0.0):\n",
    "\n",
    "\t    if first_stage_features_stride != 8 and first_stage_features_stride != 16:\n",
    "\t      raise ValueError('`first_stage_features_stride` must be 8 or 16.')\n",
    "\t    super(FasterRCNNInceptionResnetV2FeatureExtractor, self).__init__(\n",
    "\t        is_training, first_stage_features_stride, batch_norm_trainable,\n",
    "\t        reuse_weights, weight_decay)\n",
    "\t\n",
    "\t  def preprocess(self, resized_inputs):\n",
    "\n",
    "\t    return (2.0 / 255.0) * resized_inputs - 1.0\n",
    "\t\n",
    "\t  def _extract_proposal_features(self, preprocessed_inputs, scope):\n",
    "\n",
    "\t    if len(preprocessed_inputs.get_shape().as_list()) != 4:\n",
    "\t      raise ValueError('`preprocessed_inputs` must be 4 dimensional, got a '\n",
    "\t                       'tensor of shape %s' % preprocessed_inputs.get_shape())\n",
    "\t\n",
    "\t    with slim.arg_scope(inception_resnet_v2.inception_resnet_v2_arg_scope(\n",
    "\t        weight_decay=self._weight_decay)):\n",
    "\t      # Forces is_training to False to disable batch norm update.\n",
    "\t      with slim.arg_scope([slim.batch_norm],\n",
    "\t                          is_training=self._train_batch_norm):\n",
    "\t        with tf.variable_scope('InceptionResnetV2',\n",
    "\t                               reuse=self._reuse_weights) as scope:\n",
    "\t          return inception_resnet_v2.inception_resnet_v2_base(\n",
    "\t              preprocessed_inputs, final_endpoint='PreAuxLogits',\n",
    "\t              scope=scope, output_stride=self._first_stage_features_stride,\n",
    "\t              align_feature_maps=True)\n",
    "\t\n",
    "\t  def _extract_box_classifier_features(self, proposal_feature_maps, scope):\n",
    "\n",
    "\t    with tf.variable_scope('InceptionResnetV2', reuse=self._reuse_weights):\n",
    "\t      with slim.arg_scope(inception_resnet_v2.inception_resnet_v2_arg_scope(\n",
    "\t          weight_decay=self._weight_decay)):\n",
    "\t        # Forces is_training to False to disable batch norm update.\n",
    "\t        with slim.arg_scope([slim.batch_norm],\n",
    "\t                            is_training=self._train_batch_norm):\n",
    "\t          with slim.arg_scope([slim.conv2d, slim.max_pool2d, slim.avg_pool2d],\n",
    "\t                              stride=1, padding='SAME'):\n",
    "\t            with tf.variable_scope('Mixed_7a'):\n",
    "\t              with tf.variable_scope('Branch_0'):\n",
    "\t                tower_conv = slim.conv2d(proposal_feature_maps,\n",
    "\t                                         256, 1, scope='Conv2d_0a_1x1')\n",
    "\t                tower_conv_1 = slim.conv2d(\n",
    "\t                    tower_conv, 384, 3, stride=2,\n",
    "\t                    padding='VALID', scope='Conv2d_1a_3x3')\n",
    "\t              with tf.variable_scope('Branch_1'):\n",
    "\t                tower_conv1 = slim.conv2d(\n",
    "\t                    proposal_feature_maps, 256, 1, scope='Conv2d_0a_1x1')\n",
    "\t                tower_conv1_1 = slim.conv2d(\n",
    "\t                    tower_conv1, 288, 3, stride=2,\n",
    "\t                    padding='VALID', scope='Conv2d_1a_3x3')\n",
    "\t              with tf.variable_scope('Branch_2'):\n",
    "\t                tower_conv2 = slim.conv2d(\n",
    "\t                    proposal_feature_maps, 256, 1, scope='Conv2d_0a_1x1')\n",
    "\t                tower_conv2_1 = slim.conv2d(tower_conv2, 288, 3,\n",
    "\t                                            scope='Conv2d_0b_3x3')\n",
    "\t                tower_conv2_2 = slim.conv2d(\n",
    "\t                    tower_conv2_1, 320, 3, stride=2,\n",
    "\t                    padding='VALID', scope='Conv2d_1a_3x3')\n",
    "\t              with tf.variable_scope('Branch_3'):\n",
    "\t                tower_pool = slim.max_pool2d(\n",
    "\t                    proposal_feature_maps, 3, stride=2, padding='VALID',\n",
    "\t                    scope='MaxPool_1a_3x3')\n",
    "\t              net = tf.concat(\n",
    "\t                  [tower_conv_1, tower_conv1_1, tower_conv2_2, tower_pool], 3)\n",
    "\t            net = slim.repeat(net, 9, inception_resnet_v2.block8, scale=0.20)\n",
    "\t            net = inception_resnet_v2.block8(net, activation_fn=None)\n",
    "\t            proposal_classifier_features = slim.conv2d(\n",
    "\t                net, 1536, 1, scope='Conv2d_7b_1x1')\n",
    "\t        return proposal_classifier_features\n",
    "\t\n",
    "\t  def restore_from_classification_checkpoint_fn(\n",
    "\t      self,\n",
    "\t      first_stage_feature_extractor_scope,\n",
    "\t      second_stage_feature_extractor_scope):\n",
    "\n",
    "\t\n",
    "\t    variables_to_restore = {}\n",
    "\t    for variable in variables_helper.get_global_variables_safely():\n",
    "\t      if variable.op.name.startswith(\n",
    "\t          first_stage_feature_extractor_scope):\n",
    "\t        var_name = variable.op.name.replace(\n",
    "\t            first_stage_feature_extractor_scope + '/', '')\n",
    "\t        variables_to_restore[var_name] = variable\n",
    "\t      if variable.op.name.startswith(\n",
    "\t          second_stage_feature_extractor_scope):\n",
    "\t        var_name = variable.op.name.replace(\n",
    "\t            second_stage_feature_extractor_scope\n",
    "\t            + '/InceptionResnetV2/Repeat', 'InceptionResnetV2/Repeat_2')\n",
    "\t        var_name = var_name.replace(\n",
    "\t            second_stage_feature_extractor_scope + '/', '')\n",
    "\t        variables_to_restore[var_name] = variable\n",
    "\t    return variables_to_restore\n"
   ]
  }
 ],
 "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": "3.6.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}