add evonorm

ppcls/modeling/architectures/evonorm.py

@@ -0,0 +1,86 @@
+import paddle
+import torch
+import numpy as np
+import paddle.fluid as fluid
+
+def instance_std_paddle(input, epsilon=1e-5):
+    v = paddle.var(input, axis=[2,3], keepdim=True )
+    v = paddle.expand_as(v, input)
+    return  paddle.sqrt(v+epsilon)
+
+def instance_std(x, eps=1e-5):
+    var = torch.var(x, dim = (2, 3), keepdim=True).expand_as(x)
+    if torch.isnan(var).any():
+        var = torch.zeros(var.shape)
+    return torch.sqrt(var + eps)
+
+def group_std_paddle(input, groups=32, epsilon=1e-5):
+    #N, C, H, W = paddle.shape(input)
+    N,C,H,W = input.shape
+    #print(N,C,H,W)
+    input = paddle.reshape(input, [N, groups, C//groups, H, W])
+    v = paddle.var(input, axis=[2,3,4], keepdim=True)
+    v = paddle.expand_as(v, input)
+    return paddle.reshape(paddle.sqrt(v+epsilon),(N,C,H,W))
+
+def group_std(x, groups = 32, eps = 1e-5):
+    N, C, H, W = x.size()
+    x = torch.reshape(x, (N, groups, C // groups, H, W))
+    var = torch.var(x, dim = (2, 3, 4), keepdim = True).expand_as(x)
+    return torch.reshape(torch.sqrt(var + eps), (N, C, H, W))
+
+
+
+class EvoNorm(fluid.dygraph.Layer):
+    def __init__(self, channels, version='B0', affine=True, non_linear=True, groups=32, epsilon=1e-5,momentum=0.9, training=True):
+        super(EvoNorm, self).__init__()
+        self.channels = channels
+        self.affine = affine
+        self.version = version
+        self.non_linear = non_linear
+        self.groups = groups
+        self.epsilon = epsilon
+        self.training = training 
+        self.momentum = momentum
+
+        if self.affine:
+
+            self.gamma = self.create_parameter([1, self.channels, 1, 1],
+                default_initializer=fluid.initializer.Constant(value=1.0))
+            self.beta = self.create_parameter([1, self.channels, 1, 1],
+                default_initializer=fluid.initializer.Constant(value=0.0))
+            if self.non_linear:
+                self.v = self.create_parameter([1, self.channels, 1, 1],
+                    default_initializer=fluid.initializer.Constant(value=1.0))
+        else:
+            self.register_parameter('gamma', None)
+            self.register_parameter('beta', None)
+            self.register_buffer('v', None)
+
+        #self.running_var = self.create_parameter([1, self.channels, 1, 1],
+        #        default_initializer=fluid.initializer.Constant(value=0.0))
+        #self.running_var.stop_gradient = True
+        #self.register_buffer('running_var', self.create_parameter([1, self.channels, 1, 1],
+        #           default_initializer=fluid.initializer.Constant(value=1.0)))
+        self.register_buffer('running_var', paddle.fluid.layers.ones(shape=[1,self.channels,1,1], dtype='float32'))
+
+    def forward(self, input):
+        if self.version == 'S0':
+            if self.non_linear:
+                num = input * paddle.fluid.layers.sigmoid(self.v * input)
+                return num / group_std_paddle(input, groups=self.groups, epsilon=self.epsilon) * self.gamma + self.beta
+            else:
+                return input * self.gamma + self.beta
+        if self.version == 'B0':
+            if self.training:
+                var = paddle.var(input, axis=[0,2,3], unbiased=False, keepdim=True)
+                self.running_var = self.running_var * self.momentum
+                self.running_var = self.running_var + (1- self.momentum) * var
+            else:
+                var = self.running_var
+            if self.non_linear:
+                den = paddle.elementwise_max(paddle.sqrt((var+self.epsilon)), self.v * input + instance_std_paddle(input, epsilon=self.epsilon))
+                return input / den * self.gamma + self.beta
+            else:
+                return input * self.gamma + self.beta
+

ppcls/modeling/architectures/resnet.py

@@ -77,15 +77,15 @@ class BottleneckBlock(fluid.dygraph.Layer):
             filter_size=1,
             act=None)
 
-        if not shortcut:
+        self.shortcut = shortcut
+
+        if not self.shortcut:
             self.short = ConvBNLayer(
                 num_channels=num_channels,
                 num_filters=num_filters * 4,
                 filter_size=1,
                 stride=stride)
 
-        self.shortcut = shortcut
-
         self._num_channels_out = num_filters * 4
 
     def forward(self, inputs):
@@ -108,18 +108,16 @@ class ResNet(fluid.dygraph.Layer):
     def __init__(self, layers=50, class_dim=1000):
         super(ResNet, self).__init__()
 
-        self.layers = layers
-        supported_layers = [50, 101, 152]
-        assert layers in supported_layers, \
-            "supported layers are {} but input layer is {}".format(
-                supported_layers, layers)
-
-        if layers == 50:
+        if layers == 18:
+            depth = [2, 2, 2, 2]
+        elif layers == 18 or layers == 50:
             depth = [3, 4, 6, 3]
         elif layers == 101:
             depth = [3, 4, 23, 3]
         elif layers == 152:
             depth = [3, 8, 36, 3]
+        else:
+            raise ValueError('Input layer is not supported')
         num_channels = [64, 256, 512, 1024]
         num_filters = [64, 128, 256, 512]
 
@@ -191,6 +189,6 @@ def ResNet101(**kwargs):
     return model
 
 
-def ResNet152(class_dim=1000):
-    model = ResNet(layers=152, class_dim=class_dim)
+def ResNet152(**kwargs):
+    model = ResNet(layers=152, **kwargs)
     return model