Merge pull request #1209 from lamblin/micklat-fix_mlp_softmax_misclass_rebased

Fix mlp softmax misclass and cost_from_X when the top layer is Softmax with binary_target_dim

Author: lamblin

pylearn2/models/mlp.py

@@ -1131,7 +1131,7 @@ def cost_from_X_data_specs(self):
         This is useful if cost_from_X is used in a MethodCost.
         """
         space = CompositeSpace((self.get_input_space(),
-                                self.get_output_space()))
+                                self.get_target_space()))
         source = (self.get_input_source(), self.get_target_source())
         return (space, source)
 
@@ -1315,7 +1315,8 @@ def get_layer_monitoring_channels(self, state_below=None,
 
             if targets is not None:
                 y_hat = T.argmax(state, axis=1)
-                y = T.argmax(targets, axis=1)
+                y = (targets.reshape(y_hat.shape) if self._has_binary_target
+                     else T.argmax(targets, axis=1))
                 misclass = T.neq(y, y_hat).mean()
                 misclass = T.cast(misclass, config.floatX)
                 rval['misclass'] = misclass

pylearn2/models/tests/test_mlp.py

@@ -400,6 +400,46 @@ def test_softmax_binary_targets():
     np.testing.assert_allclose(cost_bin(X_data, y_bin_data),
                                cost_vec(X_data, y_vec_data))
 
+def test_softmax_bin_targets_channels(seed=0):
+    """
+    Constructs softmax layers with binary target and with vector targets
+    to check that they give the same 'misclass' channel value.
+    """
+    np.random.seed(seed)
+    num_classes = 2
+    batch_size = 5
+    mlp_bin = MLP(
+        layers=[Softmax(num_classes, 's1', irange=0.1,
+                        binary_target_dim=1)],
+        nvis=100
+    )
+    mlp_vec = MLP(
+        layers=[Softmax(num_classes, 's1', irange=0.1)],
+        nvis=100
+    )
+
+    X = mlp_bin.get_input_space().make_theano_batch()
+    y_bin = mlp_bin.get_target_space().make_theano_batch()
+    y_vec = mlp_vec.get_target_space().make_theano_batch()
+
+    X_data = np.random.random(size=(batch_size, 100))
+    X_data = X_data.astype(theano.config.floatX)
+    y_bin_data = np.random.randint(low=0, high=num_classes,
+                                   size=(batch_size, 1))
+    y_vec_data = np.zeros((batch_size, num_classes), dtype=theano.config.floatX)
+    y_vec_data[np.arange(batch_size),y_bin_data.flatten()] = 1
+
+    def channel_value(channel_name, model, y, y_data):
+        chans = model.get_monitoring_channels((X,y))
+        f_channel = theano.function([X,y], chans['s1_'+channel_name])
+        return f_channel(X_data, y_data)
+
+    for channel_name in ['misclass', 'nll']:
+        vec_val = channel_value(channel_name, mlp_vec, y_vec, y_vec_data)
+        bin_val = channel_value(channel_name, mlp_bin, y_bin, y_bin_data)
+        print channel_name, vec_val, bin_val
+        np.testing.assert_allclose(vec_val, bin_val)
+    
 def test_set_get_weights_Softmax():
     """
     Tests setting and getting weights for Softmax layer.