buf fix: sampling

Yusuke Sugomori · Yusuke Sugomori · commit 5b420c40a4a9 · 2013-01-05T21:01:00.000+09:00
diff --git a/DeepBeliefNets.py b/DeepBeliefNets.py
@@ -55,7 +55,7 @@ def __init__(self, input=None, label=None,\
             if i == 0:
                 layer_input = self.x
             else:
-                layer_input = self.sigmoid_layers[-1].output()
+                layer_input = self.sigmoid_layers[-1].sample_h_given_v()
                 
             # construct sigmoid_layer
             sigmoid_layer = HiddenLayer(input=layer_input,
@@ -76,7 +76,7 @@ def __init__(self, input=None, label=None,\
 
 
         # layer for output using Logistic Regression
-        self.log_layer = LogisticRegression(input=self.sigmoid_layers[-1].output(),
+        self.log_layer = LogisticRegression(input=self.sigmoid_layers[-1].sample_h_given_v(),
                                             label=self.y,
                                             n_in=hidden_layer_sizes[-1],
                                             n_out=n_outs)
@@ -92,9 +92,8 @@ def pretrain(self, lr=0.1, k=1, epochs=100):
             if i == 0:
                 layer_input = self.x
             else:
-                layer_input = self.sigmoid_layers[i-1].output()
+                layer_input = self.sigmoid_layers[i-1].sample_h_given_v(layer_input)
             rbm = self.rbm_layers[i]
-
             
             for epoch in xrange(epochs):
                 c = []
@@ -108,7 +107,7 @@ def pretrain(self, lr=0.1, k=1, epochs=100):
 
 
     def finetune(self, lr=0.1, epochs=100):
-        layer_input = self.sigmoid_layers[-1].output()
+        layer_input = self.sigmoid_layers[-1].sample_h_given_v()
 
         # train log_layer
         epoch = 0
diff --git a/HiddenLayer.py b/HiddenLayer.py
@@ -56,3 +56,12 @@ def output(self, input=None):
                 else self.activation(linear_output))
 
 
+    def sample_h_given_v(self, input=None):
+        if input is not None:
+            self.input = input
+
+        v_mean = self.output()
+        h_sample = self.numpy_rng.binomial(size=v_mean.shape,
+                                           n=1,
+                                           p=v_mean)
+        return h_sample
