Merge pull request #3799 from lamblin/avg_pool_v3

theano/sandbox/cuda/dnn.py

@@ -1401,6 +1401,10 @@ class GpuDnnPoolDesc(GpuOp):
             raise RuntimeError("CuDNN pooling with padding requires CuDNN v2")
         if self.get_ndim() == 3 and version() < (3000, 3000):
             raise RuntimeError("CuDNN 3d pooling requires CuDNN v3")
+        if (mode == 'average_exc_pad' and max(pad) > 0 and
+                version() < (4004, 4004)):
+            raise RuntimeError(
+                "CuDNN pooling mode 'average_exc_pad' requires at least v4")
 
     def get_ndim(self):
         return len(self.ws)

theano/sandbox/cuda/tests/test_dnn.py

@@ -235,8 +235,14 @@ def test_pooling():
     if not cuda.dnn.dnn_available():
         raise SkipTest(cuda.dnn.dnn_available.msg)
 
+    # 'average_exc_pad' is disabled for versions < 4004
+    if cuda.dnn.version() < (4004, 4004):
+        modes = ('max', 'average_inc_pad')
+    else:
+        modes = ('max', 'average_inc_pad', 'average_exc_pad')
+
     x = T.ftensor4()
-    for mode, pad in product(('max', 'average_inc_pad', 'average_exc_pad'),
+    for mode, pad in product(modes,
                              ((0, 0), (1, 0), (1, 0), (2, 3), (3, 2))):
         if mode == 'max':
             func = T.max
@@ -346,9 +352,15 @@ def test_pooling3d():
     if not cuda.dnn.dnn_available() or cuda.dnn.version() < (3000, 3000):
         raise SkipTest(cuda.dnn.dnn_available.msg)
 
+    # 'average_exc_pad' is disabled for versions < 4004
+    if cuda.dnn.version() < (4004, 4004):
+        modes = ('max', 'average_inc_pad')
+    else:
+        modes = ('max', 'average_inc_pad', 'average_exc_pad')
+
     x = T.TensorType(broadcastable=(False, False, False, False, False),
                      dtype='float32')()
-    for mode, pad in product(('max', 'average_inc_pad', 'average_exc_pad'),
+    for mode, pad in product(modes,
                              ((0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1),
                               (2, 3, 2), (3, 2, 2), (2, 2, 3))):
         if mode == 'max':
@@ -940,10 +952,17 @@ class TestDnnInferShapes(utt.InferShapeTester):
             numpy.random.rand(2, 3, 4, 5),
             dtype='float32'
         )
+
+        # 'average_exc_pad' is disabled for versions < 4004
+        if cuda.dnn.version() < (4004, 4004):
+            modes = ['max', 'average_inc_pad']
+        else:
+            modes = ['max', 'average_inc_pad', 'average_exc_pad']
+
         for params in product(
             [(1, 1), (2, 2), (3, 3)],
             [(1, 1), (2, 2), (3, 3)],
-            ['max', 'average_inc_pad', 'average_exc_pad']
+            modes
         ):
             desc = dnn.GpuDnnPoolDesc(
                 ws=params[0],

theano/sandbox/gpuarray/dnn.py

@@ -953,6 +953,9 @@ class GpuDnnPoolDesc(Op):
 
         if self.get_ndim() == 3 and version() < 3000:
             raise RuntimeError("CuDNN 3d pooling requires v3")
+        if mode == 'average_exc_pad' and max(pad) > 0 and version() < 4004:
+            raise RuntimeError(
+                "CuDNN pooling mode 'average_exc_pad' requires at least v4")
 
     def get_ndim(self):
         return len(self.ws)

theano/sandbox/gpuarray/tests/test_dnn.py

@@ -166,8 +166,14 @@ def test_pooling():
     if not dnn.dnn_available(test_ctx_name):
         raise SkipTest(dnn.dnn_available.msg)
 
+    # 'average_exc_pad' is disabled for versions < 4004
+    if dnn.version() < 4004:
+        modes = ('max', 'average_inc_pad')
+    else:
+        modes = ('max', 'average_inc_pad', 'average_exc_pad')
+
     x = T.ftensor4()
-    for mode, pad in product(('max', 'average_inc_pad', 'average_exc_pad'),
+    for mode, pad in product(modes,
                              ((0, 0), (1, 0), (1, 0), (2, 3), (3, 2))):
         if mode == 'max':
             func = T.max
@@ -506,10 +512,17 @@ class TestDnnInferShapes(utt.InferShapeTester):
             numpy.random.rand(2, 3, 4, 5),
             dtype='float32'
         )
+
+        # 'average_exc_pad' is disabled for versions < 4004
+        if dnn.version() < 4004:
+            modes = ['max', 'average_inc_pad']
+        else:
+            modes = ['max', 'average_inc_pad', 'average_exc_pad']
+
         for params in product(
             [(1, 1), (2, 2), (3, 3)],
             [(1, 1), (2, 2), (3, 3)],
-            ['max', 'average_inc_pad', 'average_exc_pad']
+            modes
         ):
             desc = dnn.GpuDnnPoolDesc(
                 ws=params[0],