Minor fixes found when running

frcnn/anchor_target.py

@@ -53,14 +53,18 @@ def __init__(self, anchors, feat_stride=[16], name='anchor_target'):
         self._batch_size = 256
         self._bbox_inside_weights = (1.0, 1.0, 1.0, 1.0)
 
-    def _build(self, rpn_cls_score, gt_boxes, im_info):
+    def _build(self, rpn_cls_prob, gt_boxes, im_info):
         """
         Args:
-            rpn_cls_score: A Tensor with the class score for every anchor
+            rpn_cls_prob: A Tensor with the class probability for every anchor
                 generated by the Region Proposal Network. It's dimensions
-                should be TODO: Chequear dimensiones.
+                should be (1, H, W, num_anchors * 2). We only use it to get the
+                pretrained shape (H, W).
             gt_boxes: A Tensor with the groundtruth bounding boxes of the
-                image of the batch being processed.
+                image of the batch being processed. It's dimensions should
+                be (num_gt, 5).
+            im_info: Shape of original image (height, width) in order to define
+                anchor targers in respect with gt_boxes.
 
         We currently use the `anchor_target_layer` code provided in the
         original Caffe implementation by Ross Girshick. Ideally we should
@@ -80,31 +84,30 @@ def _build(self, rpn_cls_score, gt_boxes, im_info):
 
         (
             labels, bbox_targets,
-            bbox_inside_weights, bbox_outside_weights
+            # bbox_inside_weights, bbox_outside_weights
         ) = tf.py_func(
             self._anchor_target_layer_np,
-            [rpn_cls_score, gt_boxes, im_info],
-            [tf.float32, tf.float32, tf.float32, tf.float32]
+            [rpn_cls_prob, gt_boxes, im_info],
+            [tf.float32, tf.float32]
 
         )
 
         return labels, bbox_targets  # missing bbox_inside_weights, bbox_outside_weights
 
 
-    def _anchor_target_layer(self, rpn_cls_score, gt_boxes, im_info):
+    def _anchor_target_layer(self, rpn_cls_prob, gt_boxes, im_info):
         """
         Function working with Tensors instead of instances for proper
         computing in the Tensorflow graph.
         """
         raise NotImplemented()
 
 
-    def _anchor_target_layer_np(self, rpn_cls_score, gt_boxes, im_info):
+    def _anchor_target_layer_np(self, rpn_cls_prob, gt_boxes, im_info):
         """
         Function to be executed with tf.py_func
         """
-
-        height, width = rpn_cls_score.shape[1:3]
+        height, width = rpn_cls_prob.shape[1:3]  # TODO(debug): rpn_cls_prob.shape = (1, 23, 279, 2)
 
         # 1. Generate proposals from bbox deltas and shifted anchors
         shift_x = np.arange(0, width) * self._feat_stride
@@ -190,8 +193,8 @@ def _anchor_target_layer_np(self, rpn_cls_score, gt_boxes, im_info):
                 bg_inds, size=(len(bg_inds) - num_bg), replace=False)
             labels[disable_inds] = -1
 
-        bbox_targets = np.zeros((len(inds_inside), 4), dtype=np.float32)
-        bbox_targets = self._compute_targets(anchors, gt_boxes[argmax_overlaps, :])
+        # TODO: Not necessary to define first bbox_targets = np.zeros((len(inds_inside), 4), dtype=np.float32)
+        bbox_targets = self._compute_targets(anchors, gt_boxes[argmax_overlaps, :]).astype(np.float32)
 
         bbox_inside_weights = np.zeros((len(inds_inside), 4), dtype=np.float32)
         bbox_inside_weights[labels == 1, :] = np.array(self._bbox_inside_weights)
@@ -211,13 +214,13 @@ def _anchor_target_layer_np(self, rpn_cls_score, gt_boxes, im_info):
         bbox_outside_weights[labels == 1, :] = positive_weights
         bbox_outside_weights[labels == 0, :] = negative_weights
 
-        labels = _unmap(
+        labels = self._unmap(
             labels, total_anchors, inds_inside, fill=-1)
-        bbox_targets = _unmap(
+        bbox_targets = self._unmap(
             bbox_targets, total_anchors, inds_inside, fill=0)
-        bbox_inside_weights = _unmap(
+        bbox_inside_weights = self._unmap(
             bbox_inside_weights, total_anchors, inds_inside, fill=0)
-        bbox_outside_weights = _unmap(
+        bbox_outside_weights = self._unmap(
             bbox_outside_weights, total_anchors, inds_inside, fill=0)
 
         # labels
@@ -239,7 +242,7 @@ def _anchor_target_layer_np(self, rpn_cls_score, gt_boxes, im_info):
             (1, height, width, A * 4)
         ).transpose(0, 3, 1, 2)
 
-        return labels, bbox_targets, bbox_inside_weights, bbox_outside_weights
+        return labels, bbox_targets
 
     def _bbox_overlaps(self, boxes, gt_boxes):
         """

frcnn/anchor_target_test.py

@@ -21,7 +21,7 @@ def setUp(self):
     def testBasic(self):
         model = AnchorTarget(self.anchors)
         rpn_cls_score_shape = (1, 32, 32, model._num_anchors * 2)
-        gt_boxes_shape = (3, 4)  # 3 ground truth boxes.
+        gt_boxes_shape = (1, 4)  # 1 ground truth boxes.
         im_info_shape = (2,)
 
         rpn_cls_score_ph = tf.placeholder(
@@ -36,8 +36,8 @@ def testBasic(self):
             sess.run(tf.global_variables_initializer())
             out_inst = sess.run(out, feed_dict={
                 rpn_cls_score_ph: np.random.rand(*rpn_cls_score_shape),
-                gt_boxes_ph: np.random.rand(*gt_boxes_shape),
-                im_info_ph: np.random.rand(*im_info_shape),
+                gt_boxes_ph: [[ 45,  42, 455, 342]],
+                im_info_ph: [375, 500],
             })
 
 

frcnn/network.py

@@ -3,7 +3,6 @@
 import tensorflow as tf
 
 from .anchor_target import AnchorTarget
-from .config import Config
 from .dataset import TFRecordDataset
 from .pretrained import VGG
 from .proposal import Proposal
@@ -52,18 +51,19 @@ def _build(self, image, gt_boxes, is_training=True):
         pretrained_output = self._pretrained(image)
         rpn_layers = self._rpn(pretrained_output, is_training=is_training)
         # TODO: We should rename the
-        rpn_cls_prob = rpn_layers['rpn_cls_prob_reshape']
-        rpn_bbox_pred = rpn_layers['rpn_bbox_pred']
+        rpn_cls_prob = rpn_layers['rpn_cls_prob_reshape']  # TODO(debug): shape (1, 21, 31, 18)
+        rpn_bbox_pred = rpn_layers['rpn_bbox_pred']  # TODO(debug): shape (1, 21, 31, 36)
 
         rpn_labels, rpn_bbox = self._anchor_target(
-            rpn_layers['rpn_cls_score_reshape'], gt_boxes, image_shape)
+            rpn_cls_prob, gt_boxes, image_shape)
 
         blob, scores = self._proposal(
             rpn_layers['rpn_cls_prob'], rpn_layers['rpn_bbox_pred'])
         roi_pool = self._roi_pool(blob, pretrained_output)
 
         # TODO: Missing mapping classification_bbox to real coordinates.
         # (and trimming, and NMS?)
+        # TODO: Missing gt_boxes labels!
         classification_prob, classification_bbox = self._rcnn(roi_pool)
 
         # TODO: We are returning only rpn tensors for training RPN.

frcnn/rpn.py

@@ -70,12 +70,12 @@ def _build(self, pretrained, is_training=True):
         """
         TODO: We don't have BatchNorm yet.
         """
-        rpn = self._rpn_activation(self._rpn(pretrained))
-        rpn_cls_score = self._rpn_cls(rpn)
-        rpn_cls_score_reshape = self.spatial_reshape_layer(rpn_cls_score, 2)
-        rpn_cls_prob = self.spatial_softmax(rpn_cls_score_reshape)
-        rpn_cls_prob_reshape = self.spatial_reshape_layer(rpn_cls_prob, self._num_anchors * 2)
-        rpn_bbox_pred = self._rpn_bbox(rpn)
+        rpn = self._rpn_activation(self._rpn(pretrained))  # TODO(debug): shape (1, 21, 31, 512)
+        rpn_cls_score = self._rpn_cls(rpn)  # TODO(debug): shape (1, 21, 31, 18)
+        rpn_cls_score_reshape = self.spatial_reshape_layer(rpn_cls_score, 2)  # TODO(debug): shape (1, 21, 279, 2)
+        rpn_cls_prob = self.spatial_softmax(rpn_cls_score_reshape)  # TODO(debug): shape (1, 21, 279, 2)
+        rpn_cls_prob_reshape = self.spatial_reshape_layer(rpn_cls_prob, self._num_anchors * 2)  # TODO(debug): shape (1, 21, 31, 18)
+        rpn_bbox_pred = self._rpn_bbox(rpn)  # TODO(debug): shape (1, 21, 31, 36)
 
         return {
             'rpn': rpn,