SimAM

model/attention/SimAM.py

@@ -0,0 +1,29 @@
+"""
+MindSpore implementation of 'SimAM'
+Refer to "SimAM: A Simple, Parameter-Free Attention Module for Convolutional Neural Networks"
+"""
+import mindspore as ms
+from mindspore import nn
+
+
+class SimAM(nn.Cell):
+    """ SimAM """
+    def __init__(self, e_lambda=1e-4):
+        super().__init__()
+        self.activation = nn.Sigmoid()
+        self.e_lambda = e_lambda
+
+    def construct(self, x):
+        _, _, H, W = x.shape
+        n = W * H - 1
+
+        x_minus_mu_square = (x - x.mean(axis=(2, 3), keep_dims=True)).pow(2)
+        y = x_minus_mu_square / (4 * x_minus_mu_square.sum(axis=(2, 3), keepdims=True) / n + self.e_lambda) + 0.5
+        return x * self.activation(y)
+
+
+if __name__ == '__main__':
+    dummy_input = ms.ops.randn(3, 64, 7, 7)
+    model = SimAM(64)
+    outputs = model(dummy_input)
+    print(outputs.shape)