dbg ghostnet

ppcls/modeling/architectures/__init__.py

@@ -27,6 +27,7 @@ from .hrnet import HRNet_W18_C
 from .efficientnet import EfficientNetB0
 from .resnest import ResNeSt50_fast_1s1x64d, ResNeSt50
 from .googlenet import GoogLeNet
+from .ghostnet import GhostNet_x0_5, GhostNet_x1_0, GhostNet_x1_3
 from .mobilenet_v1 import MobileNetV1_x0_25, MobileNetV1_x0_5, MobileNetV1_x0_75, MobileNetV1
 from .mobilenet_v2 import MobileNetV2_x0_25, MobileNetV2_x0_5, MobileNetV2_x0_75, MobileNetV2, MobileNetV2_x1_5, MobileNetV2_x2_0
 from .mobilenet_v3 import MobileNetV3_small_x0_35, MobileNetV3_small_x0_5, MobileNetV3_small_x0_75, MobileNetV3_small_x1_0, MobileNetV3_small_x1_25, MobileNetV3_large_x0_35, MobileNetV3_large_x0_5, MobileNetV3_large_x0_75, MobileNetV3_large_x1_0, MobileNetV3_large_x1_25

ppcls/modeling/architectures/ghostnet.py

@@ -20,7 +20,6 @@ import paddle.nn.functional as F
 from paddle.nn import Conv2d, BatchNorm, AdaptiveAvgPool2d, Linear
 from paddle.fluid.regularizer import L2DecayRegularizer
 from paddle.nn.initializer import Uniform
-from paddle import fluid
 
 
 class ConvBNLayer(nn.Layer):
@@ -42,9 +41,12 @@ class ConvBNLayer(nn.Layer):
             stride=stride,
             padding=(filter_size - 1) // 2,
             groups=groups,
-            weight_attr=ParamAttr(name=name + "_weights"),
-            bias_attr=False)
+            weight_attr=ParamAttr(initializer=nn.initializer.MSRA(), name=name + "_weights"),
+            bias_attr=False
+        )
         bn_name = name + "_bn"
+
+        # In the old version, moving_variance_name was name + "_variance"
         self._batch_norm = BatchNorm(
             num_filters,
             act=act,
@@ -104,7 +106,7 @@ class SEBlock(nn.Layer):
         squeeze = self.squeeze(pool)
         squeeze = F.relu(squeeze)
         excitation = self.excitation(squeeze)
-        excitation = F.sigmoid(excitation)
+        excitation = paddle.fluid.layers.clip(x=excitation, min=0, max=1)
         excitation = paddle.reshape(
             excitation,
             shape=[-1, self._num_channels, 1, 1]
@@ -138,7 +140,7 @@ class GhostModule(nn.Layer):
             name=name + "_primary_conv"
         )
         self.cheap_operation = ConvBNLayer(
-            num_channels=num_channels,
+            num_channels=init_channels,
             num_filters=new_channels,
             filter_size=dw_size,
             stride=1,
@@ -186,7 +188,7 @@ class GhostBottleneck(nn.Layer):
                 stride=stride,
                 groups=hidden_dim,
                 act=None,
-                name=name+"_depthwise"
+                name=name+"_depthwise"  # In the old version, name was name + "_depthwise_depthwise"
             )
         if use_se:
             self.se_block = SEBlock(
@@ -194,7 +196,7 @@ class GhostBottleneck(nn.Layer):
                 name=name + "_se"
             )
         self.ghost_module_2 = GhostModule(
-            num_channels=num_channels,
+            num_channels=hidden_dim,
             output_channels=output_channels,
             kernel_size=1,
             relu=False,
@@ -208,7 +210,7 @@ class GhostBottleneck(nn.Layer):
                 stride=stride,
                 groups=num_channels,
                 act=None,
-                name=name + "_shotcut_depthwise"
+                name=name + "_shortcut_depthwise"  # In the old version, name was name + "_shortcut_depthwise_depthwise"
             )
             self.shortcut_conv = ConvBNLayer(
                 num_channels=num_channels,
@@ -217,11 +219,11 @@ class GhostBottleneck(nn.Layer):
                 stride=1,
                 groups=1,
                 act=None,
-                name=name + "_shotcut_conv"
+                name=name + "_shortcut_conv"
             )
 
     def forward(self, inputs):
-        x = self.ghost_module(inputs)
+        x = self.ghost_module_1(inputs)
         if self._stride == 2:
             x = self.depthwise_conv(x)
         if self._use_se:
@@ -275,14 +277,17 @@ class GhostNet(nn.Layer):
             num_channels = output_channels
             output_channels = int(self._make_divisible(c * self.scale, 4))
             hidden_dim = int(self._make_divisible(exp_size, self.scale, 4))
-            ghost_bottleneck = GhostBottleneck(
-                num_channels=num_channels,
-                hidden_dim=hidden_dim,
-                output_channels=output_channels,
-                kernel_size=k,
-                stride=s,
-                use_se=use_se,
-                name="_ghostbottleneck" + str(idx)
+            ghost_bottleneck = self.add_sublayer(
+                name="_ghostbottleneck_" + str(idx),
+                sublayer=GhostBottleneck(
+                    num_channels=num_channels,
+                    hidden_dim=hidden_dim,
+                    output_channels=output_channels,
+                    kernel_size=k,
+                    stride=s,
+                    use_se=use_se,
+                    name="_ghostbottleneck_" + str(idx)
+                )
             )
             self.ghost_bottleneck_list.append(ghost_bottleneck)
             idx += 1
@@ -300,24 +305,26 @@ class GhostNet(nn.Layer):
         )
         self.pool2d_gap = AdaptiveAvgPool2d(1)
         num_channels = output_channels
-        output_channels = 1280
+        self._num_channels = num_channels
+        self._fc0_output_channels = 1280
         self.fc_0 = ConvBNLayer(
             num_channels=num_channels,
-            num_filters=output_channels,
+            num_filters=self._fc0_output_channels,
             filter_size=1,
             stride=1,
             act="relu",
             name="fc_0"
         )
         self.dropout = nn.Dropout(p=0.2)
-        stdv = 1.0 / math.sqrt(output_channels * 1.0)
+        stdv = 1.0 / math.sqrt(self._fc0_output_channels * 1.0)
         self.fc_1 = Linear(
-            output_channels,
+            self._fc0_output_channels,
             class_dim,
-            param_attr=ParamAttr(
+            weight_attr=ParamAttr(
                 name="fc_1_weights",
                 initializer=Uniform(-stdv, stdv)
-            )
+            ),
+            bias_attr=ParamAttr(name="fc_1_offset")
         )
 
     def forward(self, inputs):
@@ -328,6 +335,7 @@ class GhostNet(nn.Layer):
         x = self.pool2d_gap(x)
         x = self.fc_0(x)
         x = self.dropout(x)
+        x = paddle.reshape(x, shape=[-1, self._fc0_output_channels])
         x = self.fc_1(x)
         return x
 
@@ -345,3 +353,18 @@ class GhostNet(nn.Layer):
         if new_v < 0.9 * v:
             new_v += divisor
         return new_v
+
+
+def GhostNet_x0_5():
+    model = GhostNet(scale=0.5)
+    return model
+
+
+def GhostNet_x1_0():
+    model = GhostNet(scale=1.0)
+    return model
+
+
+def GhostNet_x1_3():
+    model = GhostNet(scale=1.3)
+    return model
\ No newline at end of file