fix a bug in evaluating accuracy (#2736)

* fix accuracy computation in MNIST siamese graph example

previous code:
```
def compute_accuracy(predictions, labels):
    '''Compute classification accuracy with a fixed threshold on distances.
    '''
    return labels[predictions.ravel() < 0.5].mean()
```
is not accuracy over all the samples, but over samples with negative prediction.

* add space around  "=="

follow François's suggestion

examples/mnist_siamese_graph.py

@@ -75,7 +75,7 @@ def create_base_network(input_dim):
 def compute_accuracy(predictions, labels):
     '''Compute classification accuracy with a fixed threshold on distances.
     '''
-    return labels[predictions.ravel() < 0.5].mean()
+    return np.mean(labels == (predictions.ravel() > 0.5))
 
 
 # the data, shuffled and split between train and test sets