replace F.droput with nn.Dropout (#520)

* fix a shape typo

* replace F.droput with nn.Dropout

* add ann 80k configs

* add ann 80k configs

Author: aigcliu

dygraph/configs/ann/ann_resnet101_os8_cityscapes_1024x512_160k.yml

@@ -1,6 +1,3 @@
-_base_: 'ann_resnet50_os8_cityscapes_1024x512_160k.yml'
+_base_: 'ann_resnet101_os8_cityscapes_1024x512_80k.yml'
 
-model:
-  backbone:
-    type: ResNet101_vd
-    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet101_vd_ssld.tar.gz
+iters: 160000

dygraph/configs/ann/ann_resnet101_os8_cityscapes_1024x512_80k.yml

@@ -0,0 +1,6 @@
+_base_: 'ann_resnet50_os8_cityscapes_1024x512_80k.yml'
+
+model:
+  backbone:
+    type: ResNet101_vd
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet101_vd_ssld.tar.gz

dygraph/configs/ann/ann_resnet50_os8_cityscapes_1024x512_160k.yml

@@ -1,26 +1,3 @@
-_base_: '../_base_/cityscapes.yml'
+_base_: 'ann_resnet50_os8_cityscapes_1024x512_80k.yml'
 
-batch_size: 2
 iters: 160000
-  
-learning_rate:
-  decay:
-    end_lr: 1.0e-5
-
-loss:
-  types:
-    - type: CrossEntropyLoss
-  coef: [1, 0.4]
-
-model:
-  type: ANN
-  backbone:
-    type: ResNet50_vd
-    output_stride: 8
-    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet50_vd_ssld_v2.tar.gz
-  backbone_indices: [2, 3]
-  key_value_channels: 256
-  inter_channels: 512
-  psp_size: [1, 3, 6, 8]
-  enable_auxiliary_loss: True
-  pretrained: null

dygraph/configs/ann/ann_resnet50_os8_cityscapes_1024x512_80k.yml

@@ -0,0 +1,26 @@
+_base_: '../_base_/cityscapes.yml'
+
+batch_size: 2
+iters: 80000
+
+learning_rate:
+  decay:
+    end_lr: 1.0e-5
+
+loss:
+  types:
+    - type: CrossEntropyLoss
+  coef: [1, 0.4]
+
+model:
+  type: ANN
+  backbone:
+    type: ResNet50_vd
+    output_stride: 8
+    pretrained: https://bj.bcebos.com/paddleseg/dygraph/resnet50_vd_ssld_v2.tar.gz
+  backbone_indices: [2, 3]
+  key_value_channels: 256
+  inter_channels: 512
+  psp_size: [1, 3, 6, 8]
+  enable_auxiliary_loss: True
+  pretrained: null

dygraph/paddleseg/models/ann.py

@@ -203,7 +203,7 @@ def __init__(self,
             in_channels=out_channels + high_in_channels,
             out_channels=out_channels,
             kernel_size=1)
-        self.dropout_prob = dropout_prob
+        self.dropout = nn.Dropout(p=dropout_prob)
 
     def forward(self, low_feats, high_feats):
         priors = [stage(low_feats, high_feats) for stage in self.stages]
@@ -212,7 +212,7 @@ def forward(self, low_feats, high_feats):
             context += priors[i]
 
         output = self.conv_bn(paddle.concat([context, high_feats], axis=1))
-        output = F.dropout(output, p=self.dropout_prob)  # dropout_prob
+        output = self.dropout(output)
 
         return output
 
@@ -251,7 +251,7 @@ def __init__(self,
             in_channels=in_channels * 2,
             out_channels=out_channels,
             kernel_size=1)
-        self.dropout_prob = dropout_prob
+        self.dropout = nn.Dropout(p=dropout_prob)
 
     def forward(self, x):
         priors = [stage(x) for stage in self.stages]
@@ -260,7 +260,7 @@ def forward(self, x):
             context += priors[i]
 
         output = self.conv_bn(paddle.concat([context, x], axis=1))
-        output = F.dropout(output, p=self.dropout_prob)  # dropout_prob
+        output = self.dropout(output)
 
         return output
 

dygraph/paddleseg/models/fast_scnn.py

@@ -26,14 +26,11 @@
 class FastSCNN(nn.Layer):
     """
     The FastSCNN implementation based on PaddlePaddle.
-
     As mentioned in the original paper, FastSCNN is a real-time segmentation algorithm (123.5fps)
     even for high resolution images (1024x2048).
-
     The original article refers to
     Poudel, Rudra PK, et al. "Fast-scnn: Fast semantic segmentation network"
     (https://arxiv.org/pdf/1902.04502.pdf).
-
     Args:
         num_classes (int): The unique number of target classes.
         enable_auxiliary_loss (bool, optional): A bool value indicates whether adding auxiliary loss.
@@ -95,9 +92,7 @@ def init_weight(self):
 class LearningToDownsample(nn.Layer):
     """
     Learning to downsample module.
-
     This module consists of three downsampling blocks (one conv and two separable conv)
-
     Args:
         dw_channels1 (int, optional): The input channels of the first sep conv. Default: 32.
         dw_channels2 (int, optional): The input channels of the second sep conv. Default: 48.
@@ -132,10 +127,8 @@ def forward(self, x):
 class GlobalFeatureExtractor(nn.Layer):
     """
     Global feature extractor module.
-
     This module consists of three InvertedBottleneck blocks (like inverted residual introduced by MobileNetV2) and
     a PPModule (introduced by PSPNet).
-
     Args:
         in_channels (int, optional): The number of input channels to the module. Default: 64.
         block_channels (tuple, optional): A tuple represents output channels of each bottleneck block. Default: (64, 96, 128).
@@ -189,7 +182,6 @@ def forward(self, x):
 class InvertedBottleneck(nn.Layer):
     """
     Single Inverted bottleneck implementation.
-
     Args:
         in_channels (int): The number of input channels to bottleneck block.
         out_channels (int): The number of output channels of bottleneck block.
@@ -236,9 +228,7 @@ def forward(self, x):
 class FeatureFusionModule(nn.Layer):
     """
     Feature Fusion Module Implementation.
-
     This module fuses high-resolution feature and low-resolution feature.
-
     Args:
         high_in_channels (int): The channels of high-resolution feature (output of LearningToDownsample).
         low_in_channels (int): The channels of low-resolution feature (output of GlobalFeatureExtractor).
@@ -278,9 +268,7 @@ def forward(self, high_res_input, low_res_input):
 class Classifier(nn.Layer):
     """
     The Classifier module implementation.
-
     This module consists of two depth-wise conv and one conv.
-
     Args:
         input_channels (int): The input channels to this module.
         num_classes (int): The unique number of target classes.
@@ -304,9 +292,11 @@ def __init__(self, input_channels, num_classes):
         self.conv = nn.Conv2D(
             in_channels=input_channels, out_channels=num_classes, kernel_size=1)
 
+        self.dropout = nn.Dropout(p=0.1)  # dropout_prob
+
     def forward(self, x):
         x = self.dsconv1(x)
         x = self.dsconv2(x)
-        x = F.dropout(x, p=0.1)  # dropout_prob
+        x = self.dropout(x)
         x = self.conv(x)
         return x

dygraph/paddleseg/models/gcnet.py

@@ -124,6 +124,8 @@ def __init__(self,
             kernel_size=3,
             padding=1)
 
+        self.dropout = nn.Dropout(p=0.1)
+
         self.conv = nn.Conv2D(
             in_channels=gc_channels, out_channels=num_classes, kernel_size=1)
 
@@ -147,7 +149,7 @@ def forward(self, feat_list):
         output = paddle.concat([x, output], axis=1)
         output = self.conv_bn_relu3(output)
 
-        output = F.dropout(output, p=0.1)  # dropout_prob
+        output = self.dropout(output)
         logit = self.conv(output)
         logit_list.append(logit)
 

dygraph/paddleseg/models/layers/layer_libs.py

@@ -151,15 +151,15 @@ def __init__(self,
             kernel_size=3,
             padding=1)
 
+        self.dropout = nn.Dropout(p=dropout_prob)
+
         self.conv = nn.Conv2D(
             in_channels=inter_channels,
             out_channels=out_channels,
             kernel_size=1)
 
-        self.dropout_prob = dropout_prob
-
     def forward(self, x):
         x = self.conv_bn_relu(x)
-        x = F.dropout(x, p=self.dropout_prob)
+        x = self.dropout(x)
         x = self.conv(x)
         return x

dygraph/paddleseg/models/pspnet.py

@@ -113,6 +113,8 @@ def __init__(self,
             out_channels=pp_out_channels,
             bin_sizes=bin_sizes)
 
+        self.dropout = nn.Dropout(p=0.1)  # dropout_prob
+
         self.conv = nn.Conv2D(
             in_channels=pp_out_channels,
             out_channels=num_classes,
@@ -130,7 +132,7 @@ def forward(self, feat_list):
         logit_list = []
         x = feat_list[self.backbone_indices[1]]
         x = self.psp_module(x)
-        x = F.dropout(x, p=0.1)  # dropout_prob
+        x = self.dropout(x)
         logit = self.conv(x)
         logit_list.append(logit)