未验证 提交 ba2aad26 编写于 作者: G Guanghua Yu 提交者: GitHub

fix dynamic shape of reshape op when export model (#7230)

上级 b0620a7b
......@@ -260,7 +260,7 @@ class GFLHead(nn.Layer):
center_points = paddle.stack([x, y], axis=-1)
cls_score = cls_score.reshape([b, -1, self.cls_out_channels])
bbox_pred = self.distribution_project(bbox_pred) * stride
bbox_pred = bbox_pred.reshape([b, cell_h * cell_w, 4])
bbox_pred = bbox_pred.reshape([-1, cell_h * cell_w, 4])
# NOTE: If keep_ratio=False and image shape value that
# multiples of 32, distance2bbox not set max_shapes parameter
......
......@@ -353,13 +353,13 @@ class PicoHead(OTAVFLHead):
bbox_pred = bbox_pred.reshape([1, (self.reg_max + 1) * 4,
-1]).transpose([0, 2, 1])
else:
b, _, h, w = fpn_feat.shape
_, _, h, w = fpn_feat.shape
l = h * w
cls_score_out = F.sigmoid(
cls_score.reshape([b, self.cls_out_channels, l]))
cls_score.reshape([-1, self.cls_out_channels, l]))
bbox_pred = bbox_pred.transpose([0, 2, 3, 1])
bbox_pred = self.distribution_project(bbox_pred)
bbox_pred = bbox_pred.reshape([b, l, 4])
bbox_pred = bbox_pred.reshape([-1, l, 4])
cls_logits_list.append(cls_score_out)
bboxes_reg_list.append(bbox_pred)
......@@ -597,7 +597,7 @@ class PicoHeadV2(GFLHead):
anchor_points, stride_tensor = self._generate_anchors(fpn_feats)
cls_score_list, box_list = [], []
for i, (fpn_feat, stride) in enumerate(zip(fpn_feats, self.fpn_stride)):
b, _, h, w = fpn_feat.shape
_, _, h, w = fpn_feat.shape
# task decomposition
conv_cls_feat, se_feat = self.conv_feat(fpn_feat, i)
cls_logit = self.head_cls_list[i](se_feat)
......@@ -620,10 +620,11 @@ class PicoHeadV2(GFLHead):
[0, 2, 1]))
else:
l = h * w
cls_score_out = cls_score.reshape([b, self.cls_out_channels, l])
cls_score_out = cls_score.reshape(
[-1, self.cls_out_channels, l])
bbox_pred = reg_pred.transpose([0, 2, 3, 1])
bbox_pred = self.distribution_project(bbox_pred)
bbox_pred = bbox_pred.reshape([b, l, 4])
bbox_pred = bbox_pred.reshape([-1, l, 4])
cls_score_list.append(cls_score_out)
box_list.append(bbox_pred)
......
......@@ -192,7 +192,7 @@ class PPYOLOEHead(nn.Layer):
anchor_points, stride_tensor = self._generate_anchors(feats)
cls_score_list, reg_dist_list = [], []
for i, feat in enumerate(feats):
b, _, h, w = feat.shape
_, _, h, w = feat.shape
l = h * w
avg_feat = F.adaptive_avg_pool2d(feat, (1, 1))
cls_logit = self.pred_cls[i](self.stem_cls[i](feat, avg_feat) +
......@@ -203,7 +203,7 @@ class PPYOLOEHead(nn.Layer):
reg_dist = self.proj_conv(F.softmax(reg_dist, axis=1)).squeeze(1)
# cls and reg
cls_score = F.sigmoid(cls_logit)
cls_score_list.append(cls_score.reshape([b, self.num_classes, l]))
cls_score_list.append(cls_score.reshape([-1, self.num_classes, l]))
reg_dist_list.append(reg_dist)
cls_score_list = paddle.concat(cls_score_list, axis=-1)
......@@ -238,8 +238,8 @@ class PPYOLOEHead(nn.Layer):
return loss
def _bbox_decode(self, anchor_points, pred_dist):
b, l, _ = get_static_shape(pred_dist)
pred_dist = F.softmax(pred_dist.reshape([b, l, 4, self.reg_max + 1]))
_, l, _ = get_static_shape(pred_dist)
pred_dist = F.softmax(pred_dist.reshape([-1, l, 4, self.reg_max + 1]))
pred_dist = self.proj_conv(pred_dist.transpose([0, 3, 1, 2])).squeeze(1)
return batch_distance2bbox(anchor_points, pred_dist)
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册