Fix the accuracy change problem caused by the inconsistent default be…

…havior of resize operator.
shijun18 · Oct 26, 2020 · 0c62e98 · 0c62e98
1 parent 9ca3ebd
commit 0c62e98
Show file tree

Hide file tree

Showing 23 changed files with 190 additions and 42 deletions.
diff --git a/dygraph/configs/ann/README.md b/dygraph/configs/ann/README.md
@@ -3,3 +3,12 @@
 ## Reference
 
 > Zhu, Zhen, Mengde Xu, Song Bai, Tengteng Huang, and Xiang Bai. "Asymmetric non-local neural networks for semantic segmentation." In Proceedings of the IEEE International Conference on Computer Vision, pp. 593-602. 2019.
+
+## Performance
+
+### Cityscapes
+
+| Model | Backbone | Resolution | Training Iters | mIoU | mIoU (multi-scale) | Links |
+|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
+|ANN|ResNet50_OS8|1024x512|160000|78.13%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/ann_resnet50_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/ann_resnet50_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=e5e8fb0c5d8c81558981bcf0b403af3f)|
+|ANN|ResNet101_OS8|1024x512|160000|80.25%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/ann_resnet101_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/ann_resnet101_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=7c8bf49eeb74a02f978b4050ebbea03c)|
diff --git a/dygraph/configs/deeplabv3/README.md b/dygraph/configs/deeplabv3/README.md
@@ -3,3 +3,12 @@
 ## Reference
 
 > Chen, Liang-Chieh, George Papandreou, Florian Schroff, and Hartwig Adam. "Rethinking Atrous Convolution for Semantic Image Segmentation." arXiv preprint arXiv:1706.05587 (2017).
+
+## Performance
+
+### Cityscapes
+
+| Model | Backbone | Resolution | Training Iters | mIoU | mIoU (multi-scale) | Links |
+|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
+|DeepLabV3|ResNet50_OS8|1024x512|160000|79.60%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/deeplabv3_resnet50_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/deeplabv3_resnet50_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=9dfd191d01883cbd0e5a910def16a758)|
+|DeepLabV3|ResNet101_OS8|1024x512|160000|80.05%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/deeplabv3_resnet101_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/deeplabv3_resnet101_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=67192fd2fa1f2428afc4f5cab19ecb07)|
diff --git a/dygraph/configs/deeplabv3/deeplabv3_resnet101_os8_cityscapes_1024x512_160k.yml b/dygraph/configs/deeplabv3/deeplabv3_resnet101_os8_cityscapes_1024x512_160k.yml
@@ -1,6 +1,6 @@
-_base_: 'deeplabv3_resnet50_cityscapes_1024x512_160k.yml'
+_base_: 'deeplabv3_resnet50_os8_cityscapes_1024x512_160k.yml'
 
 model:
   backbone:
     type: ResNet101_vd
-    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet101_vd_ssld.tar.gz
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet101_vd_ssld.tar.gz
diff --git a/dygraph/configs/deeplabv3p/README.md b/dygraph/configs/deeplabv3p/README.md
@@ -3,3 +3,12 @@
 ## Reference
 
 > Chen, Liang-Chieh, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. "Encoder-decoder with atrous separable convolution for semantic image segmentation." In Proceedings of the European conference on computer vision (ECCV), pp. 801-818. 2018.
+
+## Performance
+
+### Cityscapes
+
+| Model | Backbone | Resolution | Training Iters | mIoU | mIoU (multi-scale) | Links |
+|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
+|DeepLabV3P|ResNet50_OS8|1024x512|160000|80.42%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/deeplabv3p_resnet50_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/deeplabv3p_resnet50_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=bc26244ab0b0b33896435389a2b0c8cb)|
+|DeepLabV3P|ResNet101_OS8|1024x512|160000|80.82%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/deeplabv3p_resnet101_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/deeplabv3p_resnet101_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=16ec785362f84995ec9a65384f7944db)|
diff --git a/dygraph/configs/deeplabv3p/deeplabv3p_resnet101_os8_cityscapes_1024x512_160k.yml b/dygraph/configs/deeplabv3p/deeplabv3p_resnet101_os8_cityscapes_1024x512_160k.yml
@@ -1,6 +1,6 @@
-_base_: 'deeplabv3p_resnet50_cityscapes_1024x512_160k.yml'
+_base_: 'deeplabv3p_resnet50_os8_cityscapes_1024x512_160k.yml'
 
 model:
   backbone:
     type: ResNet101_vd
-    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet101_vd_ssld.tar.gz
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet101_vd_ssld.tar.gz
diff --git a/dygraph/configs/fastscnn/README.md b/dygraph/configs/fastscnn/README.md
@@ -11,4 +11,3 @@
 | Model | Backbone | Resolution | Training Iters | mIoU | mIoU (multi-scale) | Links |
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|
 |Fast SCNN|-|1024x1024|300000|69.46%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/fastscnn_cityscapes_1024x1024_300k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/fastscnn_cityscapes_1024x1024_300k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=4de7e8f13ded9a1445ebe5204214ac83)|
-
diff --git a/dygraph/configs/fcn/README.md b/dygraph/configs/fcn/README.md
@@ -9,5 +9,5 @@
 
 | Model | Backbone | Resolution | Training Iters | mIoU | mIoU (multi-scale) | Links |
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|
-|FCN|HRNet_W18|1024x512|80000|80.33%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/fcn_hrnetw18_cityscapes_1024x512_80k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/fcn_hrnetw18_cityscapes_1024x512_80k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=141ed1c7aa77474ec2a2d063713570f9)|
-|FCN|HRNet_W48|1024x512|80000|81.12%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/fcn_hrnetw48_cityscapes_1024x512_80k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/fcn_hrnetw48_cityscapes_1024x512_80k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=6f219d4b9bab266385ab6023ea097aa6)|
+|FCN|HRNet_W18|1024x512|80000|80.34%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/fcn_hrnetw18_cityscapes_1024x512_80k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/fcn_hrnetw18_cityscapes_1024x512_80k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=141ed1c7aa77474ec2a2d063713570f9)|
+|FCN|HRNet_W48|1024x512|80000|81.17%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/fcn_hrnetw48_cityscapes_1024x512_80k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/fcn_hrnetw48_cityscapes_1024x512_80k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=6f219d4b9bab266385ab6023ea097aa6)|
diff --git a/dygraph/configs/gcnet/README.md b/dygraph/configs/gcnet/README.md
@@ -3,3 +3,12 @@
 ## Reference
 
 > Cao, Yue, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. "GCNet: Non-local networks meet squeeze-excitation networks and beyond." In Proceedings of the IEEE International Conference on Computer Vision Workshops, pp. 0-0. 2019.
+
+## Performance
+
+### Cityscapes
+
+| Model | Backbone | Resolution | Training Iters | mIoU | mIoU (multi-scale) | Links |
+|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
+|GCNet|ResNet50_OS8|1024x512|160000|78.51%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/gcnet_resnet50_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/gcnet_resnet50_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=e3801edb9a6f5b33eb890f5a1ae6ed7b)|
+|GCNet|ResNet101_OS8|1024x512|160000|79.05%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/gcnet_resnet101_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/gcnet_resnet101_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=0a16394860c6a868dc400bcc84598888)|
diff --git a/dygraph/configs/pspnet/README.md b/dygraph/configs/pspnet/README.md
@@ -3,3 +3,12 @@
 ## Reference
 
 > Zhao, Hengshuang, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. "Pyramid scene parsing network." In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2881-2890. 2017.
+
+## Performance
+
+### Cityscapes
+
+| Model | Backbone | Resolution | Training Iters | mIoU | mIoU (multi-scale) | Links |
+|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
+|PSPNet|ResNet50_OS8|1024x512|160000|78.84%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/pspnet_resnet50_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/pspnet_resnet50_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=b2040a7afd0e09ec190b2edba53d2855)|
+|PSPNet|ResNet101_OS8|1024x512|160000|79.63%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/pspnet_resnet101_os8_cityscapes_1024x512_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/pspnet_resnet101_os8_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=fb3cb027983585f5dbaf7785a54d549e)|
diff --git a/dygraph/configs/unet/README.md b/dygraph/configs/unet/README.md
@@ -9,4 +9,4 @@
 
 | Model | Backbone | Resolution | Training Iters | mIoU | mIoU (multi-scale) | Links |
 |:-:|:-:|:-:|:-:|:-:|:-:|:-:|
-|UNet|-|1024x512|160000|62.18%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/unet_cityscapes_1024x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/unet_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=cbf444781f46612a30dbab5efc4d6715)|
+|UNet|-|1024x512|160000|62.20%|-|[model](https://paddleseg.bj.bcebos.com/dygraph/unet_cityscapes_1024x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/unet_cityscapes_1024x512_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=cbf444781f46612a30dbab5efc4d6715)|
diff --git a/dygraph/paddleseg/core/train.py b/dygraph/paddleseg/core/train.py
@@ -36,7 +36,12 @@ def loss_computation(logits, label, losses):
     for i in range(len(logits)):
         logit = logits[i]
         if logit.shape[-2:] != label.shape[-2:]:
-            logit = F.interpolate(logit, label.shape[-2:], mode='bilinear')
+            logit = F.interpolate(
+                logit,
+                label.shape[-2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1)
         loss_i = losses['types'][i](logit, label)
         loss += losses['coef'][i] * loss_i
     return loss

diff --git a/dygraph/paddleseg/models/ann.py b/dygraph/paddleseg/models/ann.py
@@ -69,8 +69,12 @@ def forward(self, x):
         feat_list = self.backbone(x)
         logit_list = self.head(feat_list)
         return [
-            F.interpolate(logit, x.shape[2:], mode='bilinear')
-            for logit in logit_list
+            F.interpolate(
+                logit,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1) for logit in logit_list
         ]
 
     def init_weight(self):

diff --git a/dygraph/paddleseg/models/backbones/hrnet.py b/dygraph/paddleseg/models/backbones/hrnet.py
@@ -162,9 +162,21 @@ def forward(self, x):
         st4 = self.st4(tr3)
 
         x0_h, x0_w = st4[0].shape[2:]
-        x1 = F.interpolate(st4[1], (x0_h, x0_w), mode='bilinear')
-        x2 = F.interpolate(st4[2], (x0_h, x0_w), mode='bilinear')
-        x3 = F.interpolate(st4[3], (x0_h, x0_w), mode='bilinear')
+        x1 = F.interpolate(
+            st4[1], (x0_h, x0_w),
+            mode='bilinear',
+            align_corners=True,
+            align_mode=1)
+        x2 = F.interpolate(
+            st4[2], (x0_h, x0_w),
+            mode='bilinear',
+            align_corners=True,
+            align_mode=1)
+        x3 = F.interpolate(
+            st4[3], (x0_h, x0_w),
+            mode='bilinear',
+            align_corners=True,
+            align_mode=1)
         x = paddle.concat([st4[0], x1, x2, x3], axis=1)
 
         return [x]
@@ -587,7 +599,12 @@ def forward(self, x):
                     y = self.residual_func_list[residual_func_idx](x[j])
                     residual_func_idx += 1
 
-                    y = F.interpolate(y, residual_shape, mode='bilinear')
+                    y = F.interpolate(
+                        y,
+                        residual_shape,
+                        mode='bilinear',
+                        align_corners=True,
+                        align_mode=1)
                     residual = residual + y
                 elif j < i:
                     y = x[j]

diff --git a/dygraph/paddleseg/models/bisenet.py b/dygraph/paddleseg/models/bisenet.py
@@ -71,8 +71,12 @@ def forward(self, x):
 
         logit_list = [logit, logit1, logit2, logit3, logit4]
         logit_list = [
-            F.interpolate(logit, x.shape[2:], mode='bilinear')
-            for logit in logit_list
+            F.interpolate(
+                logit,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1) for logit in logit_list
         ]
 
         return logit_list
@@ -256,13 +260,21 @@ def forward(self, dfm, sfm):
 
         sb_feat_up = self.sb_branch_up(sfm)
         sb_feat_up = F.interpolate(
-            sb_feat_up, db_feat_keep.shape[2:], mode='bilinear')
+            sb_feat_up,
+            db_feat_keep.shape[2:],
+            mode='bilinear',
+            align_corners=True,
+            align_mode=1)
         sb_feat_up = F.sigmoid(sb_feat_up)
         db_feat = db_feat_keep * sb_feat_up
 
         sb_feat = db_feat_down * sb_feat_keep
-        sb_feat = F.interpolate(sb_feat, db_feat.shape[2:], mode='bilinear')
-
+        sb_feat = F.interpolate(
+            sb_feat,
+            db_feat.shape[2:],
+            mode='bilinear',
+            align_corners=True,
+            align_mode=1)
         return self.conv(db_feat + sb_feat)
 
 

diff --git a/dygraph/paddleseg/models/danet.py b/dygraph/paddleseg/models/danet.py
@@ -56,8 +56,12 @@ def forward(self, x):
         feats = [feats[i] for i in self.backbone_indices]
         logit_list = self.head(feats)
         logit_list = [
-            F.interpolate(logit, x.shape[2:], mode='bilinear')
-            for logit in logit_list
+            F.interpolate(
+                logit,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1) for logit in logit_list
         ]
         return logit_list
 

diff --git a/dygraph/paddleseg/models/deeplab.py b/dygraph/paddleseg/models/deeplab.py
@@ -70,8 +70,12 @@ def forward(self, x):
         feat_list = self.backbone(x)
         logit_list = self.head(feat_list)
         return [
-            F.interpolate(logit, x.shape[2:], mode='bilinear')
-            for logit in logit_list
+            F.interpolate(
+                logit,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1) for logit in logit_list
         ]
 
     def init_weight(self):
@@ -159,8 +163,12 @@ def forward(self, x):
         feat_list = self.backbone(x)
         logit_list = self.head(feat_list)
         return [
-            F.interpolate(logit, x.shape[2:], mode='bilinear')
-            for logit in logit_list
+            F.interpolate(
+                logit,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1) for logit in logit_list
         ]
 
     def init_weight(self):
@@ -228,7 +236,12 @@ def __init__(self, num_classes, in_channels):
 
     def forward(self, x, low_level_feat):
         low_level_feat = self.conv_bn_relu1(low_level_feat)
-        x = F.interpolate(x, low_level_feat.shape[2:], mode='bilinear')
+        x = F.interpolate(
+            x,
+            low_level_feat.shape[2:],
+            mode='bilinear',
+            align_corners=True,
+            align_mode=1)
         x = paddle.concat([x, low_level_feat], axis=1)
         x = self.conv_bn_relu2(x)
         x = self.conv_bn_relu3(x)

diff --git a/dygraph/paddleseg/models/fast_scnn.py b/dygraph/paddleseg/models/fast_scnn.py
@@ -66,13 +66,22 @@ def forward(self, x):
         x = self.global_feature_extractor(higher_res_features)
         x = self.feature_fusion(higher_res_features, x)
         logit = self.classifier(x)
-        logit = F.interpolate(logit, x.shape[2:], mode='bilinear')
+        logit = F.interpolate(
+            logit,
+            x.shape[2:],
+            mode='bilinear',
+            align_corners=True,
+            align_mode=1)
         logit_list.append(logit)
 
         if self.enable_auxiliary_loss:
             auxiliary_logit = self.auxlayer(higher_res_features)
             auxiliary_logit = F.interpolate(
-                auxiliary_logit, x.shape[2:], mode='bilinear')
+                auxiliary_logit,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1)
             logit_list.append(auxiliary_logit)
 
         return logit_list
@@ -252,7 +261,11 @@ def __init__(self, high_in_channels, low_in_channels, out_channels):
 
     def forward(self, high_res_input, low_res_input):
         low_res_input = F.interpolate(
-            low_res_input, scale_factor=4, mode='bilinear')
+            low_res_input,
+            scale_factor=4,
+            mode='bilinear',
+            align_corners=True,
+            align_mode=1)
         low_res_input = self.dwconv(low_res_input)
         low_res_input = self.conv_low_res(low_res_input)
         high_res_input = self.conv_high_res(high_res_input)

diff --git a/dygraph/paddleseg/models/fcn.py b/dygraph/paddleseg/models/fcn.py
@@ -62,8 +62,12 @@ def forward(self, x):
         feat_list = self.backbone(x)
         logit_list = self.head(feat_list)
         return [
-            F.interpolate(logit, x.shape[2:], mode='bilinear')
-            for logit in logit_list
+            F.interpolate(
+                logit,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1) for logit in logit_list
         ]
 
     def init_weight(self):

diff --git a/dygraph/paddleseg/models/gcnet.py b/dygraph/paddleseg/models/gcnet.py
@@ -65,8 +65,12 @@ def forward(self, x):
         feat_list = self.backbone(x)
         logit_list = self.head(feat_list)
         return [
-            F.interpolate(logit, x.shape[2:], mode='bilinear')
-            for logit in logit_list
+            F.interpolate(
+                logit,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1) for logit in logit_list
         ]
 
     def init_weight(self):

diff --git a/dygraph/paddleseg/models/layers/pyramid_pool.py b/dygraph/paddleseg/models/layers/pyramid_pool.py
@@ -76,12 +76,22 @@ def forward(self, x):
         outputs = []
         for block in self.aspp_blocks:
             y = block(x)
-            y = F.interpolate(y, x.shape[2:], mode='bilinear')
+            y = F.interpolate(
+                y,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1)
             outputs.append(y)
 
         if self.image_pooling:
             img_avg = self.global_avg_pool(x)
-            img_avg = F.interpolate(img_avg, x.shape[2:], mode='bilinear')
+            img_avg = F.interpolate(
+                img_avg,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1)
             outputs.append(img_avg)
 
         x = paddle.concat(outputs, axis=1)
@@ -155,7 +165,12 @@ def forward(self, input):
         cat_layers = []
         for stage in self.stages:
             x = stage(input)
-            x = F.interpolate(x, input.shape[2:], mode='bilinear')
+            x = F.interpolate(
+                x,
+                input.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1)
             cat_layers.append(x)
         cat_layers = [input] + cat_layers[::-1]
         cat = paddle.concat(cat_layers, axis=1)

diff --git a/dygraph/paddleseg/models/ocrnet.py b/dygraph/paddleseg/models/ocrnet.py
@@ -68,8 +68,12 @@ def forward(self, x):
         feats = [feats[i] for i in self.backbone_indices]
         logit_list = self.head(feats)
         logit_list = [
-            F.interpolate(logit, x.shape[2:], mode='bilinear')
-            for logit in logit_list
+            F.interpolate(
+                logit,
+                x.shape[2:],
+                mode='bilinear',
+                align_corners=True,
+                align_mode=1) for logit in logit_list
         ]
         return logit_list
Original file line number	Diff line number	Diff line change
Expand Up		@@ -11,4 +11,3 @@
		\| Model \| Backbone \| Resolution \| Training Iters \| mIoU \| mIoU (multi-scale) \| Links \|
		\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|:-:\|
		\|Fast SCNN\|-\|1024x1024\|300000\|69.46%\|-\|[model](https://paddleseg.bj.bcebos.com/dygraph/fastscnn_cityscapes_1024x1024_300k/model.pdparams) \\| [log](https://paddleseg.bj.bcebos.com/dygraph/fastscnn_cityscapes_1024x1024_300k/train.log) \\| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=4de7e8f13ded9a1445ebe5204214ac83)\|