Fix the LogSoftmax tests in DebugMode.

theano/tensor/nnet/nnet.py

@@ -19,9 +19,9 @@ from six.moves import xrange
 import theano
 from theano import gof
 from theano import scalar
-from theano.tensor import basic as tensor
-from theano.tensor import subtensor
-from theano.tensor import opt
+from theano.tensor import basic as tensor, subtensor, opt
+from theano.tensor.type import (values_eq_approx_remove_inf,
+                                values_eq_approx_remove_nan)
 from theano.tensor.opt import copy_stack_trace
 from theano.compile import optdb
 from theano.gof import Apply
@@ -751,7 +751,9 @@ def local_logsoftmax(node):
             isinstance(node.inputs[0].owner.op, Softmax)):
         inVars = node.inputs[0].owner.inputs[0]
         new_op = LogSoftmax()
-        return [new_op(inVars)]
+        ret = new_op(inVars)
+        ret.values_eq_approx = values_eq_approx_remove_inf
+        return [ret]
 
 
 @opt.register_specialize('stabilize', 'fast_compile')
@@ -784,7 +786,9 @@ def local_logsoftmax_grad(node):
         if grads.broadcastable[1] and not sm.broadcastable[1]:
             grads = tensor.alloc(grads, grads.shape[0], sm.shape[1])
 
-        return [grads - tensor.sum(grads, axis=1, keepdims=True) * sm]
+        ret = grads - tensor.sum(grads, axis=1, keepdims=True) * sm
+        ret.values_eq_approx = values_eq_approx_remove_nan
+        return [ret]
 
 
 def softmax_graph(c):

theano/tensor/nnet/tests/test_nnet.py

@@ -190,6 +190,9 @@ class T_LogSoftmax(utt.InferShapeTester):
         utt.verify_grad(f, [numpy.random.rand(4)])
 
     def test_allclose(self):
+        m = theano.config.mode
+        m = theano.compile.get_mode(m)
+        m.check_isfinite = False
         x, y = tensor.matrices('xy')
         # regular softmax and crossentropy
         sm = tensor.nnet.softmax(x)
@@ -215,7 +218,7 @@ class T_LogSoftmax(utt.InferShapeTester):
         # now show that the two versions result in the same crossentropy cost
         # this indicates that the forward function does provide some numerical
         # stability
-        f2 = theano.function([x, y], [cm, cm2])
+        f2 = theano.function([x, y], [cm, cm2], mode=m)
         cm_, cm2_ = f2(a, b)
         utt.assert_allclose(cm_, cm2_)
 
@@ -249,6 +252,9 @@ class T_LogSoftmax(utt.InferShapeTester):
         grad and that the new operation does not explode for big inputs.
         Note that only the grad is checked.
         """
+        m = theano.config.mode
+        m = theano.compile.get_mode(m)
+        m.check_isfinite = False
         # some inputs that are large to make the gradient explode in the non
         # optimized case
         a = numpy.exp(10*numpy.random.rand(5, 10).astype(theano.config.floatX))
@@ -258,7 +264,7 @@ class T_LogSoftmax(utt.InferShapeTester):
             logsm = tensor.log(sm)
             return logsm
         # We set step to 0.1 because for big values we need a big epsilon
-        utt.verify_grad(myfunc, [a], eps=0.1)
+        utt.verify_grad(myfunc, [a], eps=0.1, mode=m)
 
 
 class T_SoftmaxGrad(utt.InferShapeTester):