hopefully final fix to scsan with gpu ( plus test)

theano/sandbox/cuda/opt.py

@@ -1214,7 +1214,12 @@ def gpuScanOptimization(node):
                     , info
                     , typeConstructor = typeConstructor
                     ).make_node(*nw_ins).outputs
-            outputs = [safe_to_cpu(x) for x in _outputs]
+            outputs = []
+            for x,y in zip(_outputs, node.outputs):
+                if isinstance(y.type, CudaNdarrayType):
+                    outputs += [x]
+                else:
+                    outputs += [safe_to_cpu(x)]
             return outputs
     return False
 

theano/scan_module/scan_utils.py

@@ -92,7 +92,7 @@ def traverse(out, x,x_copy, d):
     inner graph of scan '''
     import theano.sandbox.cuda as cuda
     if out == x:
-        d[out] = tensor.as_tensor_variable(x_copy)
+        d[out] = cuda.gpu_from_host(x_copy)
         return d
     elif out.owner is None:
         return d

theano/scan_module/tests/test_scan.py

@@ -906,6 +906,50 @@ class T_Scan(unittest.TestCase):
         theano_v = my_f()
         assert numpy.allclose( theano_v , numpy_v[5:,:])
 
+    def test_cuda_gibbs_chain(self):
+        import theano
+        from nose.plugins.skip import SkipTest
+        from theano.sandbox import cuda
+        if cuda.cuda_available == False:
+            raise SkipTest('Optional package cuda disabled')
+
+        if theano.config.mode == 'FAST_COMPILE':
+            mode = theano.compile.mode.get_mode('FAST_RUN')
+        else:
+            mode = theano.compile.mode.get_default_mode()
+        mode = mode.including('gpu','scan')
+
+
+        rng = numpy.random.RandomState(utt.fetch_seed())
+        v_vsample = numpy.array(rng.binomial(1,0.5, size=(3,20), )
+                                , dtype = 'float32')
+        vsample = theano.shared(v_vsample)
+        import theano.sandbox.rng_mrg
+        trng = theano.sandbox.rng_mrg.MRG_RandomStreams(
+                                utt.fetch_seed())
+        def f(vsample_tm1):
+            return trng.binomial(vsample_tm1.shape, n=1, p = 0.3,
+                                 dtype='float32')*vsample_tm1
+
+        theano_vsamples, updates = theano.scan(f, [], vsample, []
+                                               , n_steps = 10
+                                               , truncate_gradient=-1
+                                               , go_backwards = False
+                                               , mode = mode
+                                              )
+        gout = theano.tensor.grad(theano_vsamples[-1].sum(), vsample)
+        my_f = theano.function([], theano_vsamples[-1]
+                               , updates = updates
+                               , allow_input_downcast = True
+                               , mode = mode
+                              )
+
+        # I leave this to tested by debugmode, this test was anyway more of
+        # doest the graph compile kind of test
+        t_result = my_f()
+
+
+
 
 
     def test_gibbs_chain(self):