refactored gpu test for conv3d2d to be skipped when cuda isn't therei.

theano/sandbox/cuda/tests/test_tensor_op.py

@@ -12,6 +12,7 @@ import theano.tensor as T
 
 # Skip test if cuda_ndarray is not available.
 import theano.sandbox.cuda as cuda
+from theano.tensor.nnet.tests import test_conv3d2d
 if cuda.cuda_available == False:
     raise SkipTest('Optional package cuda disabled')
 
@@ -133,3 +134,12 @@ def test_deepcopy():
     out = f(a_v)
     assert out is not a_v
     assert numpy.allclose(numpy.asarray(a_v), numpy.asarray(out))
+
+
+def test_get_diagonal_subtensor_view():
+    test_conv3d2d.test_get_diagonal_subtensor_view(wrap=cuda.CudaNdarray)
+
+
+def test_conv3d():
+    test_conv3d2d.test_conv3d(mode=mode_with_gpu,
+                              shared=cuda.shared_constructor)

theano/tensor/nnet/tests/test_conv3d2d.py

@@ -5,57 +5,21 @@ from scipy import ndimage
 import theano
 from theano.tensor.nnet.conv3d2d import *
 import theano.tests.unittest_tools as utt
-from theano.sandbox import cuda
 
 
 if theano.config.mode == 'FAST_COMPILE':
-    mode_with_gpu = theano.compile.mode.get_mode('FAST_RUN').including('gpu')
     mode_without_gpu = theano.compile.mode.get_mode('FAST_RUN').excluding('gpu')
 else:
-    mode_with_gpu = theano.compile.mode.get_default_mode().including('gpu')
     mode_without_gpu = theano.compile.mode.get_default_mode().excluding('gpu')
 
 
-def test_get_diagonal_subtensor_view():
-
-    x = numpy.arange(20).reshape(5, 4)
-    xv01 = get_diagonal_subtensor_view(x, 0, 1)
-
-    # test that it works in 2d
-    assert numpy.all(xv01 == [[12, 9, 6, 3], [16, 13, 10, 7]])
-
-    x = numpy.arange(24).reshape(4, 3, 2)
-    xv01 = get_diagonal_subtensor_view(x, 0, 1)
-    xv02 = get_diagonal_subtensor_view(x, 0, 2)
-    xv12 = get_diagonal_subtensor_view(x, 1, 2)
-
-    #print 'x', x
-    #print 'xv01', xv01
-    #print 'xv02', xv02
-    assert numpy.all(xv01 == [
-        [[12, 13], [8, 9], [4, 5]],
-        [[18, 19], [14, 15], [10, 11]]])
-
-    assert numpy.all(xv02 == [
-        [[6, 1], [8, 3], [10, 5]],
-        [[12, 7], [14, 9], [16, 11]],
-        [[18, 13], [20, 15], [22, 17]],
-        ])
-
-    # diagonal views of each leading matrix is the same
-    # as the slices out of the diagonal view of the entire 3d tensor
-    for xi, xvi in zip(x, xv12):
-        assert numpy.all(xvi == get_diagonal_subtensor_view(xi, 0, 1))
-
-
-def test_get_diagonal_subtensor_view_gpu():
-    x = numpy.arange(20, dtype='float32').reshape(5, 4)
-    x = cuda.CudaNdarray(x)
+def test_get_diagonal_subtensor_view(wrap=lambda a: a):
+    x = numpy.arange(20).reshape(5, 4).astype('float32')
+    x = wrap(x)
     xv01 = get_diagonal_subtensor_view(x, 0, 1)
 
     # test that it works in 2d
-    assert numpy.all(numpy.asarray(xv01) ==
-                     [[12, 9, 6, 3], [16, 13, 10, 7]])
+    assert numpy.all(numpy.asarray(xv01) == [[12, 9, 6, 3], [16, 13, 10, 7]])
 
     x = numpy.arange(24).reshape(4, 3, 2)
     xv01 = get_diagonal_subtensor_view(x, 0, 1)
@@ -77,9 +41,8 @@ def test_get_diagonal_subtensor_view_gpu():
 
     # diagonal views of each leading matrix is the same
     # as the slices out of the diagonal view of the entire 3d tensor
-    for xi, xvi in zip(x, numpy.asarray(xv12)):
-        assert numpy.all(numpy.asarray(xvi) ==
-                         numpy.asarray(get_diagonal_subtensor_view(xi, 0, 1)))
+    for xi, xvi in zip(x, xv12):
+        assert numpy.all(xvi == get_diagonal_subtensor_view(xi, 0, 1))
 
 
 def pyconv3d(signals, filters):
@@ -103,7 +66,7 @@ def pyconv3d(signals, filters):
     return rval
 
 
-def test_conv3d():
+def test_conv3d(mode=mode_without_gpu, shared=theano.tensor._shared):
 
     Ns, Ts, C, Hs, Ws = 3, 10, 3, 32, 32
     Nf, Tf, C, Hf, Wf = 32, 5 , 3, 5 , 5
@@ -115,37 +78,32 @@ def test_conv3d():
     pyres = pyconv3d(signals, filters)
     print time.time() - t0
 
-    modes = [(mode_without_gpu, theano.tensor._shared)]
-    if cuda.cuda_available:
-        modes.append((mode_with_gpu, cuda.shared_constructor))
-
-    for mode, shared in modes:
-        s_signals = shared(signals)
-        s_filters = shared(filters)
-        s_output = shared(signals*0)
-
-        out = conv3d(s_signals, s_filters,
-                     signals_shape=signals.shape,
-                     filters_shape=filters.shape)
-
-        newconv3d = theano.function([], [],
-                                    updates={s_output: out},
-                                    mode=mode)
-
-        t0 = time.time()
-        newconv3d()
-        print time.time() - t0
-        utt.assert_allclose(pyres, s_output.get_value(borrow=True))
-        gsignals, gfilters = theano.grad(out.sum(), [s_signals, s_filters])
-        gnewconv3d = theano.function([], [],
-                                     updates=[(s_filters, gfilters),
-                                              (s_signals, gsignals)],
-                                     mode=mode,
-                                     name='grad')
-
-        t0 = time.time()
-        gnewconv3d()
-        print 'grad', time.time() - t0
+    s_signals = shared(signals)
+    s_filters = shared(filters)
+    s_output = shared(signals*0)
+
+    out = conv3d(s_signals, s_filters,
+                 signals_shape=signals.shape,
+                 filters_shape=filters.shape)
+
+    newconv3d = theano.function([], [],
+                                updates={s_output: out},
+                                mode=mode)
+
+    t0 = time.time()
+    newconv3d()
+    print time.time() - t0
+    utt.assert_allclose(pyres, s_output.get_value(borrow=True))
+    gsignals, gfilters = theano.grad(out.sum(), [s_signals, s_filters])
+    gnewconv3d = theano.function([], [],
+                                 updates=[(s_filters, gfilters),
+                                          (s_signals, gsignals)],
+                                 mode=mode,
+                                 name='grad')
+
+    t0 = time.time()
+    gnewconv3d()
+    print 'grad', time.time() - t0
 
     Ns, Ts, C, Hs, Ws = 3, 3, 3, 5, 5
     Nf, Tf, C, Hf, Wf = 4, 2, 3, 2, 2