Make sure to warn people that tensor_format doesn't work.

theano/sandbox/cuda/dnn.py

@@ -1778,7 +1778,7 @@ class GpuDnnSoftmaxBase(DnnBase):
     Parameters
     ----------
     tensor_format
-        Whether the data format is 'bc01' or 'b01c'.
+        Always set this to 'bc01'.
     algo
         'fast', 'accurate' or 'log' indicating whether, respectively, computations
 	should be optimized for speed, for accuracy, or if CuDNN should rather
@@ -1793,7 +1793,13 @@ class GpuDnnSoftmaxBase(DnnBase):
     __props__ = ('tensor_format', 'mode', 'algo')
 
     def __init__(self, tensor_format, algo, mode):
-        assert(tensor_format in ('bc01', 'b01c'))
+        if tensor_format != 'bc01':
+            raise ValueError(
+                "It was discovered that since December 2014, the "
+                "tensor_format parameter was ignored and the equivalent of "
+                "'bc01' is always used.  Since your code seems to be using "
+                "another value, this might have affected previous results "
+                "ran with this code.")
         DnnBase.__init__(self)
         self.tensor_format = tensor_format
 
@@ -1935,7 +1941,7 @@ class GpuDnnSoftmax(GpuDnnSoftmaxBase):
     Parameters
     ----------
     tensor_format
-        Whether the data format is 'bc01' or 'b01c'.
+        Always set to 'bc01'.
     algo
         'fast' or 'accurate' indicating whether computations should be
         optimized for speed or accuracy respectively.
@@ -2003,7 +2009,7 @@ class GpuDnnSoftmaxGrad(GpuDnnSoftmaxBase):
     Parameters
     ----------
     tensor_format
-        Whether the data format is 'bc01' or 'b01c'.
+        Always set to 'bc01'.
     algo
         'fast' or 'accurate' indicating whether computations should be
         optimized for speed or accuracy respectively.

theano/sandbox/gpuarray/dnn.py

@@ -1078,8 +1078,6 @@ class GpuDnnSoftmaxBase(DnnBase):
 
     Parameters
     ----------
-    tensor_format
-       *deprecated* Ignored, will look at the strides of the input(s).
     algo
         'fast' or 'accurate' indicating whether computations should be
         optimized for speed or accuracy respectively.
@@ -1092,7 +1090,7 @@ class GpuDnnSoftmaxBase(DnnBase):
 
     __props__ = ('mode', 'algo')
 
-    def __init__(self, _, algo, mode):
+    def __init__(self, algo, mode):
         DnnBase.__init__(self, [self.file], self.c_func)
 
         assert(algo in ('fast', 'accurate', 'log'))
@@ -1129,8 +1127,6 @@ class GpuDnnSoftmax(GpuDnnSoftmaxBase):
     """
     Op for the cuDNN Softmax.
 
-    tensor_format
-        *deprecated* Ignored, will look at input strides.
     algo
         'fast' or 'accurate' indicating whether computations should be
         optimized for speed or accuracy respectively.
@@ -1154,7 +1150,6 @@ class GpuDnnSoftmax(GpuDnnSoftmaxBase):
         g_sm, = grads
         sm = self.make_node(x).outputs[0]
         return [GpuDnnSoftmaxGrad(
-                None,
                 self.algo,
                 self.mode
                 )(g_sm, sm)]
@@ -1166,8 +1161,6 @@ class GpuDnnSoftmaxGrad(GpuDnnSoftmaxBase):
 
     Parameters
     ----------
-    tensor_format
-        *deprecated* Ignored, will look at the input strides.
     algo
         'fast' or 'accurate' indicating whether computations should be
         optimized for speed or accuracy respectively.
@@ -1393,7 +1386,7 @@ def local_softmax_dnn(node):
     if isinstance(node.op, GpuSoftmax):
         ins = node.inputs[0].dimshuffle(0, 1, 'x', 'x')
         ins = gpu_contiguous(ins)
-        out = GpuDnnSoftmax('bc01', 'accurate', 'channel')(ins)
+        out = GpuDnnSoftmax('accurate', 'channel')(ins)
         out = as_gpuarray_variable(out.dimshuffle(0, 1))
         return [out]
 
@@ -1410,7 +1403,7 @@ def local_log_softmax_dnn(node):
             isinstance(node.inputs[0].owner.op, GpuDnnSoftmax) and
             len(node.inputs[0].clients) == 1):
         softmax_node = node.inputs[0].owner
-        new_softmax = GpuDnnSoftmax(None, 'log', softmax_node.op.mode)
+        new_softmax = GpuDnnSoftmax('log', softmax_node.op.mode)
         return [new_softmax(softmax_node.inputs[0])]
 
 
@@ -1444,6 +1437,6 @@ def local_softmax_dnn_grad(node):
             return
         ins.append(n.dimshuffle(0, 1, 'x', 'x'))
 
-    out = GpuDnnSoftmaxGrad('bc01', 'accurate', 'channel')(
+    out = GpuDnnSoftmaxGrad('accurate', 'channel')(
         gpu_contiguous(ins[0]), gpu_contiguous(ins[1]))
     return [out.dimshuffle(0, 1)]

theano/sandbox/gpuarray/tests/test_dnn.py

@@ -349,7 +349,7 @@ class TestDnnInferShapes(utt.InferShapeTester):
         )
         self._compile_and_check(
             [t],
-            [dnn.GpuDnnSoftmax('bc01', 'accurate', 'channel')(t)],
+            [dnn.GpuDnnSoftmax('accurate', 'channel')(t)],
             [rand_tensor],
             dnn.GpuDnnSoftmax
         )
@@ -359,7 +359,6 @@ class TestDnnInferShapes(utt.InferShapeTester):
             [
                 T.grad(
                     dnn.GpuDnnSoftmax(
-                        'bc01',
                         'accurate',
                         'channel'
                     )(t).mean(),
@@ -707,7 +706,6 @@ class test_SoftMax(test_nnet.test_SoftMax):
         x_gpu = T.tensor4('x_gpu', 'float32')
         f_z = T.nnet.softmax_op
         f_gpu = dnn.GpuDnnSoftmax(
-            'bc01',
             'accurate',
             'channel'
         )
@@ -808,7 +806,7 @@ class test_SoftMax(test_nnet.test_SoftMax):
             raise SkipTest("Log-softmax is only in cudnn v3+")
 
         x = T.ftensor4()
-        softmax_out = dnn.GpuDnnSoftmax('bc01', 'accurate', 'channel')(x)
+        softmax_out = dnn.GpuDnnSoftmax('accurate', 'channel')(x)
         log_out = T.log(T.as_tensor_variable(softmax_out))
 
         f = theano.function([x], log_out, mode=mode_with_gpu)