fix pred (#450)

* fix pred

* fix resnest

* fix hrnet se

* fix se for export

ppcls/modeling/architectures/ghostnet.py

@@ -83,13 +83,12 @@ class SEBlock(nn.Layer):
 
     def forward(self, inputs):
         pool = self.pool2d_gap(inputs)
-        pool = paddle.reshape(pool, shape=[-1, self._num_channels])
+        pool = paddle.squeeze(pool, axis=[2, 3])
         squeeze = self.squeeze(pool)
         squeeze = F.relu(squeeze)
         excitation = self.excitation(squeeze)
         excitation = paddle.clip(x=excitation, min=0, max=1)
-        excitation = paddle.reshape(
+        excitation = paddle.unsqueeze(excitation, axis=[2, 3])
-            excitation, shape=[-1, self._num_channels, 1, 1])
         out = inputs * excitation
         return out
 

ppcls/modeling/architectures/hrnet.py

@@ -322,8 +322,7 @@ class SELayer(nn.Layer):
         self.squeeze = Linear(
             num_channels,
             med_ch,
-            act="relu",
+            weight_attr=ParamAttr(
-            param_attr=ParamAttr(
                 initializer=Uniform(-stdv, stdv), name=name + "_sqz_weights"),
             bias_attr=ParamAttr(name=name + '_sqz_offset'))
 
@@ -331,18 +330,18 @@ class SELayer(nn.Layer):
         self.excitation = Linear(
             med_ch,
             num_filters,
-            act="sigmoid",
+            weight_attr=ParamAttr(
-            param_attr=ParamAttr(
                 initializer=Uniform(-stdv, stdv), name=name + "_exc_weights"),
             bias_attr=ParamAttr(name=name + '_exc_offset'))
 
     def forward(self, input):
         pool = self.pool2d_gap(input)
-        pool = paddle.reshape(pool, shape=[-1, self._num_channels])
+        pool = paddle.squeeze(pool, axis=[2, 3])
         squeeze = self.squeeze(pool)
+        squeeze = F.relu(squeeze)
         excitation = self.excitation(squeeze)
-        excitation = paddle.reshape(
+        excitation = F.sigmoid(excitation)
-            excitation, shape=[-1, self._num_channels, 1, 1])
+        excitation = paddle.unsqueeze(excitation, axis=[2, 3])
         out = input * excitation
         return out
 

ppcls/modeling/architectures/resnest.py

@@ -85,7 +85,8 @@ class rSoftmax(nn.Layer):
             x = paddle.reshape(
                 x=x,
                 shape=[
-                    batch, cardinality, radix, int(r * h * w / cardinality / radix)
+                    batch, cardinality, radix,
+                    int(r * h * w / cardinality / radix)
                 ])
             x = paddle.transpose(x=x, perm=[0, 2, 1, 3])
             x = nn.functional.softmax(x, axis=1)
@@ -164,7 +165,7 @@ class SplatConv(nn.Layer):
 
         atten = self.conv3(gap)
         atten = self.rsoftmax(atten)
-        atten = paddle.reshape(x=atten, shape=[-1, atten.shape[1], 1, 1])
+        atten = paddle.unsqueeze(x=atten, axis=[2, 3])
 
         if self.radix > 1:
             attens = paddle.split(atten, num_or_sections=self.radix, axis=1)

ppcls/modeling/architectures/se_resnet_vd.py

@@ -223,13 +223,12 @@ class SELayer(nn.Layer):
 
     def forward(self, input):
         pool = self.pool2d_gap(input)
-        pool = paddle.reshape(pool, shape=[-1, self._num_channels])
+        pool = paddle.squeeze(pool, axis=[2, 3])
         squeeze = self.squeeze(pool)
         squeeze = F.relu(squeeze)
         excitation = self.excitation(squeeze)
         excitation = F.sigmoid(excitation)
-        excitation = paddle.reshape(
+        excitation = paddle.unsqueeze(excitation, axis=[2, 3])
-            excitation, shape=[-1, self._num_channels, 1, 1])
         out = input * excitation
         return out
 

ppcls/modeling/architectures/se_resnext.py

@@ -157,13 +157,12 @@ class SELayer(nn.Layer):
 
     def forward(self, input):
         pool = self.pool2d_gap(input)
-        pool = paddle.reshape(pool, shape=[-1, self._num_channels])
+        pool = paddle.squeeze(pool, axis=[2, 3])
         squeeze = self.squeeze(pool)
         squeeze = self.relu(squeeze)
         excitation = self.excitation(squeeze)
         excitation = self.sigmoid(excitation)
-        excitation = paddle.reshape(
+        excitation = paddle.unsqueeze(excitation, axis=[2, 3])
-            excitation, shape=[-1, self._num_channels, 1, 1])
         out = input * excitation
         return out
 

ppcls/modeling/architectures/se_resnext_vd.py

@@ -164,13 +164,12 @@ class SELayer(nn.Layer):
 
     def forward(self, input):
         pool = self.pool2d_gap(input)
-        pool = paddle.reshape(pool, shape=[-1, self._num_channels])
+        pool = paddle.squeeze(pool, axis=[2, 3])
         squeeze = self.squeeze(pool)
         squeeze = self.relu(squeeze)
         excitation = self.excitation(squeeze)
         excitation = self.sigmoid(excitation)
-        excitation = paddle.reshape(
+        excitation = paddle.unsqueeze(excitation, axis=[2, 3])
-            excitation, shape=[-1, self._num_channels, 1, 1])
         out = input * excitation
         return out
 

tools/infer/predict.py

@@ -90,7 +90,11 @@ def main(args):
         predictor.run()
 
         output = output_tensor.copy_to_cpu()
-        return utils.postprocess(output, args)
+        classes, scores = utils.postprocess(output, args)
+        print("Current image file: {}".format(args.image_file))
+        print("\ttop-1 class: {0}".format(classes[0]))
+        print("\ttop-1 score: {0}".format(scores[0]))
+
     else:
         for i in range(0, test_num + 10):
             inputs = np.random.rand(args.batch_size, 3, 224,
@@ -115,7 +119,4 @@ def main(args):
 
 if __name__ == "__main__":
     args = utils.parse_args()
-    classes, scores = main(args)
+    main(args)
-    print("Current image file: {}".format(args.image_file))
-    print("\ttop-1 class: {0}".format(classes[0]))
-    print("\ttop-1 score: {0}".format(scores[0]))