Explicitly calls .sum() to get a scalar cost before computing gradients.

theano/tensor/tests/mlp_test.py

@@ -278,10 +278,11 @@ def test_mlp():
     classifier = MLP( rng = rng, input=x, n_in=28*28, n_hidden = 500, n_out=10)
 
     # the cost we minimize during training is the negative log likelihood of 
-    # the model
-    cost = classifier.negative_log_likelihood(y)
+    # the model.
+    # We take the mean of the cost over each minibatch.
+    cost = classifier.negative_log_likelihood(y).mean()
 
-    # compute the gradient of cost with respect to theta (sotred in params)
+    # compute the gradient of cost with respect to theta (stored in params)
     # the resulting gradients will be stored in a list gparams
     gparams = []
     for param in classifier.params:

theano/tensor/tests/test_basic.py

@@ -926,7 +926,7 @@ class T_max_and_argmax(unittest.TestCase):
 
         utt.verify_grad(lambda v: max_and_argmax(v,axis=[0])[0], [data])
         utt.verify_grad(lambda v: max_and_argmax(v,axis=[0])[1], [data])
-        check_grad_max(data,eval_outputs(grad(max_and_argmax(n,axis=0)[0],n)),axis=0)
+        check_grad_max(data,eval_outputs(grad(max_and_argmax(n,axis=0)[0].sum(),n)),axis=0)
 
         utt.verify_grad(lambda v: max_and_argmax(v,axis=[1])[0], [data])
         utt.verify_grad(lambda v: max_and_argmax(v,axis=[1])[1], [data])
@@ -1241,7 +1241,7 @@ class T_min_max(unittest.TestCase):
         utt.verify_grad(lambda v: max(v), [data])
 
         utt.verify_grad(lambda v: max(v,axis=[0]), [data])
-        check_grad_max(data,eval_outputs(grad(max(n,axis=0),n)),axis=0)
+        check_grad_max(data,eval_outputs(grad(max(n,axis=0).sum(),n)),axis=0)
 
         utt.verify_grad(lambda v: max(v,axis=[1]), [data])
         #check_grad_max(data,eval_outputs(grad(max(n,axis=1),n)),axis=1)
@@ -1273,7 +1273,7 @@ class T_min_max(unittest.TestCase):
         utt.verify_grad(lambda v: min(v), [data])
 
         utt.verify_grad(lambda v: min(v,axis=[0]), [data])
-        check_grad_min(data,eval_outputs(grad(min(n,axis=0),n)),axis=0)
+        check_grad_min(data,eval_outputs(grad(min(n,axis=0).sum(),n)),axis=0)
 
         utt.verify_grad(lambda v: min(v,axis=[1]), [data])
         #check_grad_min(data,eval_outputs(grad(min(n,axis=1),n)),axis=1)

theano/tensor/tests/test_opt.py

@@ -1843,7 +1843,7 @@ class T_local_erfc(unittest.TestCase):
         mode_fusion = copy.copy(self.mode_fusion)
         mode_fusion.check_isfinite = False
 
-        f = theano.function([x],T.grad(T.log(T.erfc(x)),x), mode=mode)
+        f = theano.function([x],T.grad(T.log(T.erfc(x)).sum(),x), mode=mode)
         #theano.printing.debugprint(f)
         assert len(f.maker.env.nodes)==23, len(f.maker.env.nodes)
         assert all(numpy.isfinite(f(val)))
@@ -1878,13 +1878,13 @@ class T_local_erfc(unittest.TestCase):
         assert all(numpy.isfinite(f(val)))
 
         #test that it work correctly if x is x*2 in the graph.
-        f = theano.function([x],T.grad(T.log(T.erfc(2*x)),x), mode=mode)
+        f = theano.function([x],T.grad(T.log(T.erfc(2*x)).sum(),x), mode=mode)
         #theano.printing.debugprint(f)
         assert len(f.maker.env.nodes)==23, len(f.maker.env.nodes)
         assert numpy.isfinite(f(val)).all()
         assert f.maker.env.outputs[0].dtype==theano.config.floatX
 
-        f = theano.function([x],T.grad(T.log(T.erfc(x)),x), mode=mode_fusion)
+        f = theano.function([x],T.grad(T.log(T.erfc(x)).sum(),x), mode=mode_fusion)
         assert len(f.maker.env.nodes)==1, len(f.maker.env.nodes)
         assert f.maker.env.outputs[0].dtype==theano.config.floatX