Sigmoid is now optional in the TwoLayerMLP

bob/learn/pytorch/architectures/TwoLayerMLP.py

@@ -14,10 +14,10 @@ import torch.nn.functional as F
 #==============================================================================
 # Define the network:
 
-
 class TwoLayerMLP(nn.Module):
     """
-    A simple two-layer MLP for binary classification.
+    A simple two-layer MLP for binary classification. The output activation
+    function is sigmoid.
 
     Attributes
     ----------
@@ -26,9 +26,13 @@ class TwoLayerMLP(nn.Module):
 
     n_hidden_relu : int
         Number of ReLU units in the hidden layer of the MLP.
+
+    apply_sigmoid : bool
+        If set to ``True`` the sigmoid will be applied to the output of the
+        hidden FC layer. If ``False`` the sigmoid is not applied.
     """
 
-    def __init__(self, in_features, n_hidden_relu):
+    def __init__(self, in_features, n_hidden_relu, apply_sigmoid = True):
         super(TwoLayerMLP, self).__init__()
         """
         Init method.
@@ -38,10 +42,13 @@ class TwoLayerMLP(nn.Module):
 
         self.n_hidden_relu = n_hidden_relu
 
+        self.apply_sigmoid = apply_sigmoid
+
         self.fc1 = nn.Linear(in_features = self.in_features, out_features = self.n_hidden_relu, bias=True)
 
         self.fc2 = nn.Linear(in_features = self.n_hidden_relu, out_features = 1, bias=True)
 
+
     def forward(self, x):
         """
         The forward method.
@@ -57,6 +64,11 @@ class TwoLayerMLP(nn.Module):
 
         # second fully connected activated by sigmoid:
         x = self.fc2(x)
+
+        if not self.apply_sigmoid:
+
+            return x
+
         x = F.sigmoid(x)
 
         return x