[dev] reorganize code of s2anet_head (#3481)

* reorganize the code, and fix bug

* fix typo

* remve comment

* reorganize code

* Revert "reorganize code"

This reverts commit 4f928c3a46b264ba8815978b914bf81a46b66b62.

* set weight

* set default config

* fix code style

configs/dota/_base_/s2anet.yml

@@ -36,6 +36,8 @@ S2ANetHead:
   align_conv_type: 'Conv'  # AlignConv Conv
   align_conv_size: 3
   use_sigmoid_cls: True
+  reg_loss_weight: [ 1.0, 1.0, 1.0, 1.0, 1.1 ]
+  cls_loss_weight: [ 1.1, 1.05 ]
 
 RBoxAssigner:
   pos_iou_thr: 0.5
@@ -52,4 +54,3 @@ S2ANetBBoxPostProcess:
     score_threshold: 0.05
     nms_threshold: 0.1
     normalized: False
-    #background_label: -1

configs/dota/s2anet_1x_dota.yml

-_BASE_: [
+it _BASE_: [
   '../datasets/dota.yml',
   '../runtime.yml',
   '_base_/s2anet_optimizer_1x.yml',
@@ -6,3 +6,18 @@ _BASE_: [
   '_base_/s2anet_reader.yml',
 ]
 weights: output/s2anet_1x_dota/model_final
+
+S2ANetHead:
+  anchor_strides: [8, 16, 32, 64, 128]
+  anchor_scales: [4]
+  anchor_ratios: [1.0]
+  anchor_assign: RBoxAssigner
+  stacked_convs: 2
+  feat_in: 256
+  feat_out: 256
+  num_classes: 15
+  align_conv_type: 'AlignConv'  # AlignConv Conv
+  align_conv_size: 3
+  use_sigmoid_cls: True
+  reg_loss_weight: [1.0, 1.0, 1.0, 1.0, 1.1]
+  cls_loss_weight: [1.1, 1.05]

configs/dota/s2anet_conv_1x_dota.yml

@@ -19,3 +19,5 @@ S2ANetHead:
   align_conv_type: 'Conv'  # AlignConv Conv
   align_conv_size: 3
   use_sigmoid_cls: True
+  reg_loss_weight: [1.0, 1.0, 1.0, 1.0, 1.1]
+  cls_loss_weight: [1.1, 1.05]

ppdet/modeling/bbox_utils.py

@@ -267,6 +267,150 @@ def bbox_iou(box1, box2, giou=False, diou=False, ciou=False, eps=1e-9):
         return iou
 
 
+def rect2rbox(bboxes):
+    """
+    :param bboxes: shape (n, 4) (xmin, ymin, xmax, ymax)
+    :return: dbboxes: shape (n, 5) (x_ctr, y_ctr, w, h, angle)
+    """
+    bboxes = bboxes.reshape(-1, 4)
+    num_boxes = bboxes.shape[0]
+
+    x_ctr = (bboxes[:, 2] + bboxes[:, 0]) / 2.0
+    y_ctr = (bboxes[:, 3] + bboxes[:, 1]) / 2.0
+    edges1 = np.abs(bboxes[:, 2] - bboxes[:, 0])
+    edges2 = np.abs(bboxes[:, 3] - bboxes[:, 1])
+    angles = np.zeros([num_boxes], dtype=bboxes.dtype)
+
+    inds = edges1 < edges2
+
+    rboxes = np.stack((x_ctr, y_ctr, edges1, edges2, angles), axis=1)
+    rboxes[inds, 2] = edges2[inds]
+    rboxes[inds, 3] = edges1[inds]
+    rboxes[inds, 4] = np.pi / 2.0
+    return rboxes
+
+
+def delta2rbox(rrois,
+               deltas,
+               means=[0, 0, 0, 0, 0],
+               stds=[1, 1, 1, 1, 1],
+               wh_ratio_clip=1e-6):
+    """
+    :param rrois: (cx, cy, w, h, theta)
+    :param deltas: (dx, dy, dw, dh, dtheta)
+    :param means:
+    :param stds:
+    :param wh_ratio_clip:
+    :return:
+    """
+    means = paddle.to_tensor(means)
+    stds = paddle.to_tensor(stds)
+    deltas = paddle.reshape(deltas, [-1, deltas.shape[-1]])
+    denorm_deltas = deltas * stds + means
+
+    dx = denorm_deltas[:, 0]
+    dy = denorm_deltas[:, 1]
+    dw = denorm_deltas[:, 2]
+    dh = denorm_deltas[:, 3]
+    dangle = denorm_deltas[:, 4]
+
+    max_ratio = np.abs(np.log(wh_ratio_clip))
+    dw = paddle.clip(dw, min=-max_ratio, max=max_ratio)
+    dh = paddle.clip(dh, min=-max_ratio, max=max_ratio)
+
+    rroi_x = rrois[:, 0]
+    rroi_y = rrois[:, 1]
+    rroi_w = rrois[:, 2]
+    rroi_h = rrois[:, 3]
+    rroi_angle = rrois[:, 4]
+
+    gx = dx * rroi_w * paddle.cos(rroi_angle) - dy * rroi_h * paddle.sin(
+        rroi_angle) + rroi_x
+    gy = dx * rroi_w * paddle.sin(rroi_angle) + dy * rroi_h * paddle.cos(
+        rroi_angle) + rroi_y
+    gw = rroi_w * dw.exp()
+    gh = rroi_h * dh.exp()
+    ga = np.pi * dangle + rroi_angle
+    ga = (ga + np.pi / 4) % np.pi - np.pi / 4
+    ga = paddle.to_tensor(ga)
+
+    gw = paddle.to_tensor(gw, dtype='float32')
+    gh = paddle.to_tensor(gh, dtype='float32')
+    bboxes = paddle.stack([gx, gy, gw, gh, ga], axis=-1)
+    return bboxes
+
+
+def rbox2delta(proposals, gt, means=[0, 0, 0, 0, 0], stds=[1, 1, 1, 1, 1]):
+    """
+
+    Args:
+        proposals:
+        gt:
+        means: 1x5
+        stds: 1x5
+
+    Returns:
+
+    """
+    proposals = proposals.astype(np.float64)
+
+    PI = np.pi
+
+    gt_widths = gt[..., 2]
+    gt_heights = gt[..., 3]
+    gt_angle = gt[..., 4]
+
+    proposals_widths = proposals[..., 2]
+    proposals_heights = proposals[..., 3]
+    proposals_angle = proposals[..., 4]
+
+    coord = gt[..., 0:2] - proposals[..., 0:2]
+    dx = (np.cos(proposals[..., 4]) * coord[..., 0] + np.sin(proposals[..., 4])
+          * coord[..., 1]) / proposals_widths
+    dy = (-np.sin(proposals[..., 4]) * coord[..., 0] + np.cos(proposals[..., 4])
+          * coord[..., 1]) / proposals_heights
+    dw = np.log(gt_widths / proposals_widths)
+    dh = np.log(gt_heights / proposals_heights)
+    da = (gt_angle - proposals_angle)
+
+    da = (da + PI / 4) % PI - PI / 4
+    da /= PI
+
+    deltas = np.stack([dx, dy, dw, dh, da], axis=-1)
+    means = np.array(means, dtype=deltas.dtype)
+    stds = np.array(stds, dtype=deltas.dtype)
+    deltas = (deltas - means) / stds
+    deltas = deltas.astype(np.float32)
+    return deltas
+
+
+def bbox_decode(bbox_preds,
+                anchors,
+                means=[0, 0, 0, 0, 0],
+                stds=[1, 1, 1, 1, 1]):
+    """decode bbox from deltas
+    Args:
+        bbox_preds: [N,H,W,5]
+        anchors: [H*W,5]
+    return:
+        bboxes: [N,H,W,5]
+    """
+    means = paddle.to_tensor(means)
+    stds = paddle.to_tensor(stds)
+    num_imgs, H, W, _ = bbox_preds.shape
+    bboxes_list = []
+    for img_id in range(num_imgs):
+        bbox_pred = bbox_preds[img_id]
+        # bbox_pred.shape=[5,H,W]
+        bbox_delta = bbox_pred
+        anchors = paddle.to_tensor(anchors)
+        bboxes = delta2rbox(
+            anchors, bbox_delta, means, stds, wh_ratio_clip=1e-6)
+        bboxes = paddle.reshape(bboxes, [H, W, 5])
+        bboxes_list.append(bboxes)
+    return paddle.stack(bboxes_list, axis=0)
+
+
 def poly2rbox(polys):
     """
     poly:[x0,y0,x1,y1,x2,y2,x3,y3]

ppdet/modeling/proposal_generator/target_layer.py

@@ -296,7 +296,7 @@ class RBoxAssigner(object):
             anchors = anchors.reshape(-1, anchors.shape[-1])
         assert anchors.ndim == 2
         anchor_num = anchors.shape[0]
-        anchor_valid = np.ones((anchor_num), np.uint8)
+        anchor_valid = np.ones((anchor_num), np.int32)
         anchor_inds = np.arange(anchor_num)
         return anchor_inds
 
@@ -371,9 +371,8 @@ class RBoxAssigner(object):
         # calc rbox iou
         anchors_xc_yc = anchors_xc_yc.astype(np.float32)
         gt_bboxes_xc_yc = gt_bboxes_xc_yc.astype(np.float32)
-        anchors_xc_yc = paddle.to_tensor(anchors_xc_yc, place=paddle.CPUPlace())
-        gt_bboxes_xc_yc = paddle.to_tensor(
-            gt_bboxes_xc_yc, place=paddle.CPUPlace())
+        anchors_xc_yc = paddle.to_tensor(anchors_xc_yc)
+        gt_bboxes_xc_yc = paddle.to_tensor(gt_bboxes_xc_yc)
 
         try:
             from rbox_iou_ops import rbox_iou
@@ -433,8 +432,7 @@ class RBoxAssigner(object):
         ignore_iof_thr = self.ignore_iof_thr
 
         anchor_num = anchors.shape[0]
-        anchors_inds = self.anchor_valid(anchors)
-        anchors = anchors[anchors_inds]
+
         gt_bboxes = gt_bboxes
         is_crowd_slice = is_crowd
         not_crowd_inds = np.where(is_crowd_slice == 0)