fix bug

kitti.yaml

@@ -42,37 +42,42 @@ network:
   p3_RPN_FEAT_STRIDE: 4
 
   p3_ANCHOR_RATIOS:
+  - 0.25
   - 0.5
   - 1
   - 2
+  - 3
   p3_ANCHOR_SCALES:
   - 8
-  p3_NUM_ANCHORS: 3
+  p3_NUM_ANCHORS: 5
 
 
   p4_RCNN_FEAT_STRIDE: 8
   p4_RPN_FEAT_STRIDE: 8
 
   p4_ANCHOR_RATIOS:
+  - 0.25
   - 0.5
   - 1
   - 2
+  - 3
   p4_ANCHOR_SCALES:
   - 8
-  p4_NUM_ANCHORS: 3
+  p4_NUM_ANCHORS: 5
 
 
   p5_RCNN_FEAT_STRIDE: 16
   p5_RPN_FEAT_STRIDE: 16
 
   p5_ANCHOR_RATIOS:
+  - 0.25
   - 0.5
   - 1
   - 2
+  - 3
   p5_ANCHOR_SCALES:
   - 8
-  p5_NUM_ANCHORS: 3
-
+  p5_NUM_ANCHORS: 5
 
 dataset:
   NUM_CLASSES: 4
@@ -84,14 +89,14 @@ dataset:
   proposal: rpn
 TRAIN:
   lr: 0.0001
-  lr_step: '5.333'
+  lr_step: '50.333'
 
   warmup: false
   warmup_lr: 0.02
   # typically we will use 8000 warmup step for single GPU for COCO
   warmup_step: 1000
   begin_epoch: 0
-  end_epoch: 40
+  end_epoch: 100
   model_prefix: 'rcnn_kitti'
   # whether resume training
   RESUME: false

lib/rpn/rpn.py

@@ -84,11 +84,11 @@ def assign_anchor(feat_shape_p3,feat_shape_p4,feat_shape_p5,
     'bbox_inside_weight': *todo* mark the assigned anchors
     'bbox_outside_weight': used to normalize the bbox_loss, all weights sums to RPN_POSITIVE_WEIGHT
     """
-    allowed_border=10
+
     feat_shape = [feat_shape_p3,feat_shape_p4,feat_shape_p5]
     feat_stride=[4,8,16]
     scales=(8,)
-    ratios=(0.5, 1, 2)
+    ratios=(0.25,0.5, 1, 2,3)
 
 
     def _unmap(data, count, inds, fill=0):
@@ -138,14 +138,14 @@ def _unmap(data, count, inds, fill=0):
 
          # keep only inside anchors
         anchors = all_anchors
-        inds_inside = np.where((all_anchors[:, 0] >= -allowed_border) &
-                           (all_anchors[:, 1] >= -allowed_border) &
-                           (all_anchors[:, 2] < im_info[1] + allowed_border) &
-                           (all_anchors[:, 3] < im_info[0] + allowed_border))[0]
+        # inds_inside = np.where((all_anchors[:, 0] >= -allowed_border) &
+        #                    (all_anchors[:, 1] >= -allowed_border) &
+        #                    (all_anchors[:, 2] < im_info[1] + allowed_border) &
+        #                    (all_anchors[:, 3] < im_info[0] + allowed_border))[0]
         # label: 1 is positive, 0 is negative, -1 is dont care
-        total_anchors = len(inds_inside)#3*w*h
-        anchors = all_anchors[inds_inside, :]
-        labels = np.empty((len(inds_inside),), dtype=np.float32)
+        total_anchors = len(anchors)#3*w*h
+     #   anchors = all_anchors[inds_inside, :]
+        labels = np.empty((total_anchors,), dtype=np.float32)
         labels.fill(-1)
 
         if gt_boxes.size > 0:
@@ -185,9 +185,9 @@ def _unmap(data, count, inds, fill=0):
 
 
         # map up to original set of anchors
-        labels = _unmap(labels, int(K * A), inds_inside, fill=-1)
-        bbox_targets = _unmap(bbox_targets, int(K * A), inds_inside, fill=0)
-        bbox_weights = _unmap(bbox_weights, int(K * A), inds_inside, fill=0)
+        labels = _unmap(labels, int(K * A), range(total_anchors), fill=-1)
+        bbox_targets = _unmap(bbox_targets, int(K * A), range(total_anchors), fill=0)
+        bbox_weights = _unmap(bbox_weights, int(K * A), range(total_anchors), fill=0)
 
 
 

retinanet/core/metric.py

@@ -57,6 +57,7 @@ def update(self, labels, preds):
      #   keep_inds = np.where(label != -1)
         pred_label = pred_label[keep_inds]
         label = label[keep_inds]
+
         self.sum_metric += np.sum(pred_label.flat == label.flat)
         self.num_inst += len(pred_label.flat)
 

retinanet/operator_py/focal_loss.py

@@ -24,7 +24,7 @@ def forward(self, is_train, req, in_data, out_data, aux):
 
         self._labels = labels
         pro_ = (mx.nd.sigmoid(cls_score) + 1e-14).asnumpy()
-    #    print "pro:", pro_[0]
+        print "pro:", pro_[0]
 
         # pro_ = np.exp(1e-14+cls_score - cls_score.max(axis=1).reshape((cls_score.shape[0], 1)))
         # pro_ /= pro_.sum(axis=1).reshape((cls_score.shape[0], 1))
@@ -49,7 +49,7 @@ def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
         pt = self._pt + 1e-14
 
         pt = pt.reshape(len(pt),1)
-        dx =   (1-self._alpha)*np.power(1 - pt, self._gamma - 1) * (self._gamma * (-1 * pt * pro_) * np.log(pt) + pro_ * (1 - pt)) * 1.0 
+        dx = (1-self._alpha)*np.power(1 - pt, self._gamma - 1) * (self._gamma * (-1 * pt * pro_) * np.log(pt) + pro_ * (1 - pt)) * 1.0 
 
         ####i==j 
         #reload pt
@@ -58,12 +58,13 @@ def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
         negative_inds = np.where(labels>-100)
 
         if self.use_ignore:
-
             ig_inds= np.where(labels==self.ignore_label)
             dx[ig_inds,:] =0
             negative_inds = np.where(labels!=self.ignore_label)
             pt = pt[negative_inds]
-        dx[negative_inds, labels[negative_inds].astype('int')]  = (self._alpha)*(-1)*dx[negative_inds, labels[negative_inds].astype('int')]
+        dx[negative_inds, labels[negative_inds].astype('int')]  = (self._alpha)*np.power(1 - pt, self._gamma) * (self._gamma * pt * np.log(pt) + pt -1) * (1.0)
+        #0.25*(dx[negative_inds, labels[negative_inds].astype('int')]-1)
+        # (self._alpha)*(-1)*dx[negative_inds, labels[negative_inds].astype('int')]
         #(self._alpha)*np.power(1 - pt, self._gamma) * (self._gamma * pt * np.log(pt) + pt -1) * (1.0)
 
         # print 'label:',labels[negative_inds][0]

retinanet/symbols/retina_resnet_101.py

@@ -1027,7 +1027,7 @@ def get_symbol(self, cfg, is_train=True):
         return group
 
     def init_weight(self, cfg, arg_params, aux_params):
-        pi = 0.01
+        pi = 0.00001
 
 
         arg_params['cls_conv1_3x3_weight'] = mx.random.normal(0, 0.01, shape=self.arg_shape_dict['cls_conv1_3x3_weight'])
@@ -1054,6 +1054,7 @@ def init_weight(self, cfg, arg_params, aux_params):
         arg_params['box_conv4_bias'] = mx.nd.zeros(shape=self.arg_shape_dict['box_conv4_bias'])
         arg_params['box_pred_weight'] = mx.random.normal(0, 0.01, shape=self.arg_shape_dict['box_pred_weight'])
         arg_params['box_pred_bias'] = mx.nd.zeros(shape=self.arg_shape_dict['box_pred_bias'])
+        #*(-np.log((1-pi)/pi))