Fix some tests in DebugMode. Shapes where cause nan, division by zeros. precison…

theano/gpuarray/tests/test_dnn.py

@@ -1451,7 +1451,7 @@ def test_dnn_batchnorm_train():
                                               bn.AbstractBatchNormTrainGrad)) for n
                             in f_abstract.maker.fgraph.toposort()])
             # run
-            for data_shape in ((5, 10, 30, 4, 10, 5), (4, 3, 1, 1, 1, 1), (1, 1, 5, 5, 5, 5)):
+            for data_shape in ((5, 10, 30, 4, 10, 5), (4, 3, 1, 1, 1, 1), (2, 3, 5, 5, 5, 5)):
                 data_shape = data_shape[:ndim]
                 param_shape = tuple(1 if d in axes else s
                                     for d, s in enumerate(data_shape))

theano/sandbox/cuda/dnn.py

@@ -3705,6 +3705,19 @@ def local_gpu_batch_norm_inference_inplace(node):
                                          inplace=True)(*node.inputs)]
 
 
+def values_eq_approx_high_tol(a, b):
+    """
+    This fct is needed to don't have DebugMode raise useless
+    error due to ronding error.
+
+    This happen as We reduce on the two last dimensions, so this
+    can raise the absolute error if the number of element we
+    reduce on is significant.
+
+    """
+    return tensor.TensorType.values_eq_approx(a, b, atol=0.015)
+
+
 @local_optimizer([bn.AbstractBatchNormTrainGrad])
 def local_abstract_batch_norm_train_grad_cudnn(node):
     if not isinstance(node.op, bn.AbstractBatchNormTrainGrad):
@@ -3781,6 +3794,9 @@ def local_abstract_batch_norm_train_grad_cudnn(node):
     if isinstance(node.outputs[2].type, tensor.TensorType):
         g_wrt_bias = tensor.as_tensor_variable(g_wrt_bias)
     # TODO copy_stack_trace?
+
+    g_wrt_inputs.tag.values_eq_approx = values_eq_approx_high_tol
+    g_wrt_scale.tag.values_eq_approx = values_eq_approx_high_tol
     return [g_wrt_inputs, g_wrt_scale, g_wrt_bias]
 
 

theano/sandbox/cuda/tests/test_dnn.py

@@ -801,7 +801,7 @@ def test_batchnorm_train():
                                               bn.AbstractBatchNormTrainGrad)) for n
                             in f_abstract.maker.fgraph.toposort()])
             # run
-            for data_shape in ((5, 2, 30, 4, 10, 5), (4, 3, 1, 1, 1, 1), (1, 1, 5, 5, 5, 5)):
+            for data_shape in ((5, 2, 30, 4, 10, 5), (4, 3, 1, 1, 1, 1), (2, 3, 5, 5, 5, 5)):
                 data_shape = data_shape[:ndim]
                 param_shape = tuple(1 if d in axes else s
                                     for d, s in enumerate(data_shape))
@@ -980,7 +980,7 @@ def test_batchnorm_inference():
                                               bn.AbstractBatchNormTrainGrad)) for n
                             in f_abstract.maker.fgraph.toposort()])
             # run
-            for data_shape in ((10, 2, 30, 4, 10, 5), (4, 3, 1, 1, 1, 1), (1, 1, 5, 5, 5, 5)):
+            for data_shape in ((10, 2, 15, 4, 6, 5), (4, 3, 1, 1, 1, 1), (1, 1, 5, 5, 5, 5)):
                 data_shape = data_shape[:ndim]
                 param_shape = tuple(1 if d in axes else s
                                     for d, s in enumerate(data_shape))

theano/tensor/nnet/tests/test_bn.py

@@ -201,7 +201,7 @@ def test_batch_normalization_train():
                                               bn.AbstractBatchNormTrainGrad))
                             for n in f.maker.fgraph.toposort()])
             # run
-            for data_shape in ((5, 10, 30, 40, 10), (4, 3, 1, 1, 1), (1, 1, 5, 5, 5)):
+            for data_shape in ((5, 10, 30, 40, 10), (4, 3, 1, 1, 1), (2, 3, 5, 5, 5)):
                 data_shape = data_shape[:ndim]
                 param_shape = tuple(1 if d in axes2 else s
                                     for d, s in enumerate(data_shape))