Merge pull request #6604 from abergeron/dnn_mix

theano/gpuarray/dnn.py

@@ -947,10 +947,14 @@ def _dnn_conv(img, kerns, alpha=1, beta=0, out=None, border_mode='valid', subsam
               conv_mode='conv', algo=None, precision=None, num_groups=1):
     ctx_name = infer_context_name(img, kerns)
 
-    img = gpu_contiguous(as_gpuarray_variable(img, ctx_name))
-    kerns = gpu_contiguous(as_gpuarray_variable(kerns, ctx_name))
+    img = as_gpuarray_variable(img, ctx_name)
+    kerns = as_gpuarray_variable(kerns, ctx_name)
 
     precision = get_precision(precision, [img, kerns])
+
+    img = gpu_contiguous(img.astype(precision))
+    kerns = gpu_contiguous(kerns.astype(precision))
+
     desc = GpuDnnConvDesc(border_mode=border_mode, subsample=subsample, dilation=dilation,
                           conv_mode=conv_mode, precision=precision, num_groups=num_groups)(kerns.shape)
     desc_op = desc.owner.op
@@ -974,11 +978,15 @@ def _dnn_gradweight(img, topgrad, kerns_shp, alpha=1, beta=0, out=None, border_m
                     dilation=(1, 1), conv_mode='conv', algo=None, precision=None, num_groups=1):
     ctx_name = infer_context_name(img, topgrad)
 
-    img = gpu_contiguous(as_gpuarray_variable(img, ctx_name))
-    topgrad = gpu_contiguous(as_gpuarray_variable(topgrad, ctx_name))
+    img = as_gpuarray_variable(img, ctx_name)
+    topgrad = as_gpuarray_variable(topgrad, ctx_name)
     kerns_shp = theano.tensor.as_tensor_variable(kerns_shp)
 
     precision = get_precision(precision, [img, topgrad], for_grad=True)
+
+    img = gpu_contiguous(img.astype(precision))
+    topgrad = gpu_contiguous(topgrad.astype(precision))
+
     desc = GpuDnnConvDesc(border_mode=border_mode, subsample=subsample, dilation=dilation,
                           conv_mode=conv_mode, precision=precision, num_groups=num_groups)(kerns_shp)
     if beta == 0:
@@ -995,11 +1003,15 @@ def _dnn_gradinput(kerns, topgrad, img_shp, alpha=1, beta=0, out=None, border_mo
                    dilation=(1, 1), conv_mode='conv', algo=None, precision=None, num_groups=1):
     ctx_name = infer_context_name(kerns, topgrad)
 
-    kerns = gpu_contiguous(as_gpuarray_variable(kerns, ctx_name))
-    topgrad = gpu_contiguous(as_gpuarray_variable(topgrad, ctx_name))
+    kerns = as_gpuarray_variable(kerns, ctx_name)
+    topgrad = as_gpuarray_variable(topgrad, ctx_name)
     img_shp = theano.tensor.as_tensor_variable(img_shp)
 
     precision = get_precision(precision, [kerns, topgrad], for_grad=True)
+
+    kerns = gpu_contiguous(kerns.astype(precision))
+    topgrad = gpu_contiguous(topgrad.astype(precision))
+
     desc = GpuDnnConvDesc(border_mode=border_mode, subsample=subsample, dilation=dilation,
                           conv_mode=conv_mode, precision=precision, num_groups=num_groups)(kerns.shape)
     if beta == 0:

theano/gpuarray/tests/test_dnn.py

@@ -220,6 +220,48 @@ def test_dnn_conv_invalid_precision():
     yield (run_dnn_conv_invalid_precision, 3)
 
 
+def test_dnn_conv_mixed_dtype():
+    mf = T.ftensor4()
+    md = T.dtensor4()
+
+    def assert_types(conv):
+        dt = conv.owner.inputs[0].dtype
+        assert conv.owner.inputs[1].dtype == dt
+        assert conv.owner.inputs[2].dtype == dt
+
+    assert_types(dnn.dnn_conv(md, mf, precision='as_input'))
+    assert_types(dnn.dnn_conv(mf, md, precision='as_input'))
+    assert_types(dnn.dnn_gradweight(mf, md, kerns_shp=mf.shape,
+                                    precision='as_input'))
+    assert_types(dnn.dnn_gradweight(md, mf, kerns_shp=mf.shape,
+                                    precision='as_input'))
+    assert_types(dnn.dnn_gradinput(mf, md, img_shp=mf.shape,
+                                   precision='as_input'))
+    assert_types(dnn.dnn_gradinput(md, mf, img_shp=mf.shape,
+                                   precision='as_input'))
+
+
+def test_dnn_conv3d_mixed_dtype():
+    mf = T.ftensor5()
+    md = T.dtensor5()
+
+    def assert_types(conv):
+        dt = conv.owner.inputs[0].dtype
+        assert conv.owner.inputs[1].dtype == dt
+        assert conv.owner.inputs[2].dtype == dt
+
+    assert_types(dnn.dnn_conv3d(md, mf, precision='as_input'))
+    assert_types(dnn.dnn_conv3d(mf, md, precision='as_input'))
+    assert_types(dnn.dnn_gradweight3d(mf, md, kerns_shp=mf.shape,
+                                      precision='as_input'))
+    assert_types(dnn.dnn_gradweight3d(md, mf, kerns_shp=mf.shape,
+                                      precision='as_input'))
+    assert_types(dnn.dnn_gradinput3d(mf, md, img_shp=mf.shape,
+                                     precision='as_input'))
+    assert_types(dnn.dnn_gradinput3d(md, mf, img_shp=mf.shape,
+                                     precision='as_input'))
+
+
 def test_pooling():
     if not dnn.dnn_available(test_ctx_name):
         raise SkipTest(dnn.dnn_available.msg)