def _network(self):
''' Define the network outputs '''
with tf.variable_scope('model'):
self.cnn = convnet(self.x, [5, 3, 3, 3, 3],
[64, 96, 128, 172, 256],
strides=[2, 2, 2, 2, 2])
featureMaps = self.cnn.get_output()
_feat_stride = self.cnn.get_feat_stride()
# Region Proposal Network (RPN)
self.rpn_net = rpn(featureMaps, self.gt_boxes, self.im_dims,
_feat_stride, flags)
rpn_cls_score = self.rpn_net.get_rpn_cls_score()
rpn_bbox_pred = self.rpn_net.get_rpn_bbox_pred()
# Roi Pooling
roi_proposal_net = roi_proposal(rpn_cls_score, rpn_bbox_pred,
self.gt_boxes, self.im_dims, flags)
# R-CNN Classification
self.fast_rcnn_net = fast_rcnn(featureMaps, roi_proposal_net)
with tf.variable_scope('model', reuse=True):
assert tf.get_variable_scope().reuse is True
self.cnn_valid = convnet(self.x_valid, [5, 3, 3, 3, 3],
[64, 96, 128, 172, 256],
strides=[2, 2, 2, 2, 2])
featureMaps_valid = self.cnn_valid.get_output()
_feat_stride_valid = self.cnn_valid.get_feat_stride()
# Region Proposal Network (RPN)
self.rpn_net_valid = rpn(featureMaps_valid, self.gt_boxes_valid,
self.im_dims_valid, _feat_stride_valid,
flags)
rpn_cls_score_valid = self.rpn_net_valid.get_rpn_cls_score()
rpn_bbox_pred_valid = self.rpn_net_valid.get_rpn_bbox_pred()
# Roi Pooling
self.roi_proposal_net_valid = roi_proposal(rpn_cls_score_valid,
rpn_bbox_pred_valid,
self.gt_boxes_valid,
self.im_dims_valid,
flags)
# R-CNN Classification
self.fast_rcnn_net_valid = fast_rcnn(featureMaps_valid,
self.roi_proposal_net_valid)
def _faster_rcnn(self, x, gt_boxes, im_dims, key):
self.cnn[key] = convnet(x, [5, 3, 3, 3, 3], [64, 96, 128, 172, 256], strides=[2, 2, 2, 2, 2])
featureMaps = self.cnn[key].get_output()
_feat_stride = self.cnn[key].get_feat_stride()
# Region Proposal Network (RPN)
self.rpn_net[key] = rpn(featureMaps, gt_boxes, im_dims, _feat_stride, flags)
# Roi Pooling
self.roi_proposal_net[key] = roi_proposal(self.rpn_net[key], gt_boxes, im_dims, key, flags)
# R-CNN Classification
self.fast_rcnn_net[key] = fast_rcnn(featureMaps, self.roi_proposal_net[key])
def _network(self):
''' Define the network outputs '''
with tf.variable_scope('model'):
self.cnn = convnet(self.x, [5, 3, 3, 3, 3], [64, 96, 128, 172, 256], strides=[2, 2, 2, 2, 2])
featureMaps = self.cnn.get_output()
_feat_stride = self.cnn.get_feat_stride()
# Region Proposal Network (RPN)
self.rpn_net = rpn(featureMaps, self.gt_boxes, self.im_dims, _feat_stride, flags)
rpn_cls_score = self.rpn_net.get_rpn_cls_score()
rpn_bbox_pred = self.rpn_net.get_rpn_bbox_pred()
# Roi Pooling
roi_proposal_net = roi_proposal(rpn_cls_score, rpn_bbox_pred, self.gt_boxes, self.im_dims, flags)
# R-CNN Classification
self.fast_rcnn_net = fast_rcnn(featureMaps, roi_proposal_net)
with tf.variable_scope('model', reuse=True):
assert tf.get_variable_scope().reuse is True
self.cnn_valid = convnet(self.x_valid, [5, 3, 3, 3, 3], [64, 96, 128, 172, 256], strides=[2, 2, 2, 2, 2])
featureMaps_valid = self.cnn_valid.get_output()
_feat_stride_valid = self.cnn_valid.get_feat_stride()
# Region Proposal Network (RPN)
self.rpn_net_valid = rpn(featureMaps_valid, self.gt_boxes_valid, self.im_dims_valid, _feat_stride_valid, flags)
rpn_cls_score_valid = self.rpn_net_valid.get_rpn_cls_score()
rpn_bbox_pred_valid = self.rpn_net_valid.get_rpn_bbox_pred()
# Roi Pooling
self.roi_proposal_net_valid = roi_proposal(rpn_cls_score_valid, rpn_bbox_pred_valid, self.gt_boxes_valid, self.im_dims_valid, flags)
# R-CNN Classification
self.fast_rcnn_net_valid = fast_rcnn(featureMaps_valid, self.roi_proposal_net_valid)
def _faster_rcnn(self, x, gt_boxes, im_dims, key):
# VALID and TEST are both evaluation mode
eval_mode = True if (key == 'EVAL') else False
self.cnn[key] = convnet(x, [5, 3, 3, 3, 3], [32, 64, 64, 128, 128], strides=[2, 2, 1, 2, 1])
feature_maps = self.cnn[key].get_output()
_feat_stride = self.cnn[key].get_feat_stride()
# Region Proposal Network (RPN)
self.rpn_net[key] = rpn(feature_maps, gt_boxes, im_dims, _feat_stride, eval_mode)
# RoI Proposals
self.roi_proposal_net[key] = roi_proposal(self.rpn_net[key], gt_boxes, im_dims, eval_mode)
# Fast R-CNN Classification
self.fast_rcnn_net[key] = fast_rcnn(feature_maps, self.roi_proposal_net[key], eval_mode)
def _faster_rcnn(self, x, gt_boxes, im_dims, key):
# VALID and TEST are both evaluation mode
eval_mode = True if (key == 'EVAL') else False
self.cnn[key] = convnet(x, [5, 3, 3, 3, 3], [32, 64, 64, 128, 128], strides=[2, 2, 1, 2, 1])
feature_maps = self.cnn[key].get_output()
_feat_stride = self.cnn[key].get_feat_stride()
# Region Proposal Network (RPN)
self.rpn_net[key] = rpn(feature_maps, gt_boxes, im_dims, _feat_stride, eval_mode)
# RoI Proposals
self.roi_proposal_net[key] = roi_proposal(self.rpn_net[key], gt_boxes, im_dims, eval_mode)
# Fast R-CNN Classification
self.fast_rcnn_net[key] = fast_rcnn(feature_maps, self.roi_proposal_net[key], eval_mode)
def _faster_rcnn(self, x, gt_boxes, im_dims, key):
# VALID and TEST are both evaluation mode
eval_mode = True if (key == 'EVAL') else False
# CNN Feature extractor
feature_maps = resnet50_reduced(x)
# CNN downsampling factor
_feat_stride = 16
self.print_log(feature_maps)
# Region Proposal Network (RPN)
self.rpn_net[key] = rpn(feature_maps, gt_boxes, im_dims, _feat_stride, eval_mode)
# RoI Proposals
self.roi_proposal_net[key] = roi_proposal(self.rpn_net[key], gt_boxes, im_dims, eval_mode)
# Fast R-CNN Classification
self.fast_rcnn_net[key] = fast_rcnn(feature_maps, self.roi_proposal_net[key], eval_mode)
def _faster_rcnn(self, x, gt_boxes, im_dims, key):
# VALID and TEST are both evaluation mode
eval_mode = True if (key == 'EVAL') else False
# CNN Feature extractor
feature_maps = resnet50_reduced(x)
# CNN downsampling factor
_feat_stride = 16
self.print_log(feature_maps)
# Region Proposal Network (RPN)
self.rpn_net[key] = rpn(feature_maps, gt_boxes, im_dims, _feat_stride, eval_mode)
# RoI Proposals
self.roi_proposal_net[key] = roi_proposal(self.rpn_net[key], gt_boxes, im_dims, eval_mode)
# Fast R-CNN Classification
self.fast_rcnn_net[key] = fast_rcnn(feature_maps, self.roi_proposal_net[key], eval_mode)
def _network(self):
''' Define the network outputs '''
# Convolutional Feature Extractor: ResNet101
# self.cnn = resnet(101, self.x)
# Simpler convnet for debugging
self.cnn = convnet(self.x, [5, 3, 3, 3, 3], [64, 96, 128, 172, 256],
strides=[2, 2, 2, 2, 2])
featureMaps = self.cnn.get_output()
_feat_stride = self.cnn.get_feat_stride()
# Region Proposal Network (RPN)
self.rpn_net = rpn(featureMaps, self.gt_boxes, self.im_dims,
_feat_stride, flags)
rpn_cls_score = self.rpn_net.get_rpn_cls_score()
rpn_bbox_pred = self.rpn_net.get_rpn_bbox_pred()
# ROI proposal
self.roi_proposal_net = roi_proposal(rpn_cls_score, rpn_bbox_pred,
self.gt_boxes, self.im_dims,
flags)
# R-CNN Classification
self.fast_rcnn_net = fast_rcnn(featureMaps, self.roi_proposal_net)