Merge pull request #2639 from nouiz/dnn_v2

theano/sandbox/cuda/dnn.py

@@ -7,15 +7,12 @@ from theano.scalar import as_scalar, constant
 from theano.gradient import DisconnectedType, grad_not_implemented
 from theano.gof import Optimizer, local_optimizer, COp
 from theano.gof.type import CDataType, Generic
-from theano.compat import PY3
 from theano.compile import optdb
 from theano.compile.ops import shape_i
 from theano.configparser import AddConfigVar, EnumStr
 from theano.tensor.nnet import SoftmaxGrad
 from theano.tensor.signal.downsample import (
     DownsampleFactorMax, DownsampleFactorMaxGrad)
-from theano.tensor.basic import ShapeError
-from theano.sandbox.cuda.type import CudaNdarrayType
 from theano.sandbox.cuda import GpuOp
 from theano.sandbox.cuda.basic_ops import (as_cuda_ndarray_variable,
                                            host_from_gpu,
@@ -82,6 +79,14 @@ if ((err = cudnnCreate(&_handle)) != CUDNN_STATUS_SUCCESS) {
                                          " from one version, but we link with"
                                          " a different version %s" % str(v))
                     raise RuntimeError(dnn_available.msg)
+                if version() == (20, 20):
+                    dnn_available.avail = False
+                    dnn_available.msg = (
+                        "You have installed a release candidate of CuDNN v2."
+                        " This isn't supported anymore."
+                        " Update to CuDNN v2 final version.")
+                    raise RuntimeError(dnn_available.msg)
+
     return dnn_available.avail
 
 
@@ -481,7 +486,6 @@ class GpuDnnConv(DnnBase, COp):
             (w + 2*padw - kw)//sw + 1
         )
 
-
     def infer_shape(self, node, shape):
         return [shape[2]]
 
@@ -659,7 +663,7 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
       work with this Op.
     """
     fgraph = getattr(img, 'fgraph', None) or getattr(kerns, 'fgraph', None)
-    if (border_mode == 'valid' and subsample == (1,1) and
+    if (border_mode == 'valid' and subsample == (1, 1) and
         direction_hint == 'bprop weights'):
         # Special case: We are asked to use GpuDnnConvGradW. We need to set
         # up a suitable 'fake' convolution to compute the gradient for.
@@ -702,12 +706,6 @@ def dnn_conv(img, kerns, border_mode='valid', subsample=(1, 1),
     desc = GpuDnnConvDesc(border_mode=border_mode, subsample=subsample,
                           conv_mode=conv_mode)(img.shape, kerns.shape)
     desc_op = desc.owner.op
-    if conv_mode == 'cross' and subsample != (1, 1) and border_mode != 'valid':
-        # there is a bug in cudnn v2 rc1-3 which gives incorrect
-        # results in this case when using the workmem='small'
-        # algorithm.
-        if workmem is None or workmem == 'small':
-            workmem = 'none'
     out_shp = GpuDnnConv.get_out_shape(img.shape, kerns.shape,
                                        desc_op.border_mode,
                                        desc_op.subsample)
@@ -814,8 +812,7 @@ class GpuDnnPoolDesc(GpuOp):
 """ % dict(name=name, desc=desc, mode_flag=mode_flag, fail=sub['fail'],
            wsX=self.ws[0], wsY=self.ws[1],
            stridex=self.stride[0], stridey=self.stride[1],
-           padX=self.pad[0], padY=self.pad[1],
-       )
+           padX=self.pad[0], padY=self.pad[1])
 
     def c_code_cache_version(self):
         return (2, version())
@@ -1480,7 +1477,7 @@ err%(name)s = cudnnSoftmaxBackward(
 
 # Intentation for history
 if True:
-    #@register_opt('cudnn')  # this optimizer is registered in opt.py instead.
+    # @register_opt('cudnn')  # this optimizer is registered in opt.py instead.
     @local_optimizer([GpuConv])
     def local_conv_dnn(node):
         if not dnn_available():
@@ -1535,19 +1532,19 @@ if True:
 
     @local_optimizer([GpuDnnConv], inplace=True)
     def local_dnn_conv_inplace(node):
-        if type(node.op) != GpuDnnConv or node.op.inplace == True:
+        if type(node.op) != GpuDnnConv or node.op.inplace:
             return
         return [GpuDnnConv(workmem=node.op.workmem, inplace=True)(*node.inputs)]
 
     @local_optimizer([GpuDnnConvGradW], inplace=True)
     def local_dnn_convgw_inplace(node):
-        if type(node.op) != GpuDnnConvGradW or node.op.inplace == True:
+        if type(node.op) != GpuDnnConvGradW or node.op.inplace:
             return
         return [GpuDnnConvGradW(inplace=True)(*node.inputs)]
 
     @local_optimizer([GpuDnnConvGradI], inplace=True)
     def local_dnn_convgi_inplace(node):
-        if type(node.op) != GpuDnnConvGradI or node.op.inplace == True:
+        if type(node.op) != GpuDnnConvGradI or node.op.inplace:
             return
         return [GpuDnnConvGradI(inplace=True)(*node.inputs)]
 
@@ -1656,8 +1653,8 @@ if True:
 
             if ((inp.owner and isinstance(inp.owner.op, HostFromGpu)) or
                 (out.owner and isinstance(out.owner.op, HostFromGpu)) or
-                (inp_grad.owner and isinstance(inp_grad.owner.op, HostFromGpu))
-            ):
+                (inp_grad.owner and isinstance(inp_grad.owner.op,
+                                               HostFromGpu))):
                 desc = GpuDnnPoolDesc(ws=ds, stride=st, mode="max", pad=pad)()
                 if not node.op.ignore_border:
                     return

theano/sandbox/cuda/tests/test_dnn.py

@@ -44,11 +44,11 @@ def pool_2d_i2n(input, ds=(2, 2), strides=None,
         assert pool_function is T.max
         pad_x = pad[0]
         pad_y = pad[1]
-        a = T.alloc(-numpy.inf, shape[0], shape[1], shape[2] + pad_x*2,
-                    shape[3] + pad_y*2)
+        a = T.alloc(-numpy.inf, shape[0], shape[1], shape[2] + pad_x * 2,
+                    shape[3] + pad_y * 2)
         input = T.set_subtensor(a[:, :,
-                                  pad_x:pad_x+shape[2],
-                                  pad_y:pad_y+shape[3]],
+                                  pad_x:pad_x + shape[2],
+                                  pad_y:pad_y + shape[3]],
                                 input)
         shape = input.shape
 
@@ -119,7 +119,7 @@ def test_pooling():
         # Test the grad
         for shp in [(1, 1, 2, 2),
                     (1, 1, 3, 3)]:
-            data = numpy.random.normal(0, 1, shp).astype("float32")*10
+            data = numpy.random.normal(0, 1, shp).astype("float32") * 10
 
             ws = 2
             stride = 2
@@ -214,12 +214,13 @@ def test_dnn_tag():
             [x],
             max_pool_2d(x, ds=(2, 2), ignore_border=True),
             mode=mode_with_gpu.including("cudnn"))
-    except (AssertionError, RuntimeError), e:
+    except (AssertionError, RuntimeError):
         assert not cuda.dnn.dnn_available()
         raised = True
     finally:
         theano.config.on_opt_error = old
-        logging.getLogger('theano.compile.tests.test_dnn').removeHandler(handler)
+        logging.getLogger(
+            'theano.compile.tests.test_dnn').removeHandler(handler)
         logging.getLogger('theano').addHandler(theano.logging_default_handler)
 
     if not raised:
@@ -332,7 +333,7 @@ class TestDnnInferShapes(utt.InferShapeTester):
                 kern_vals.shape[1], img_val.shape[1],
                 img_val.shape[2] - kern_vals.shape[2] + 1,
                 img_val.shape[3] - kern_vals.shape[3] + 1
-                )
+            )
             out_vals = numpy.zeros(shape, dtype='float32')
             desc = dnn.GpuDnnConvDesc(
                 border_mode=params[0],
@@ -481,7 +482,8 @@ def test_dnn_conv_merge():
     kw = 6
     img_val = numpy.random.random((b, c, ih, iw)).astype('float32')
     kern_val = numpy.random.random((f, c, kh, kw)).astype('float32')
-    out_val = numpy.random.random((b, f, ih-kh+1, iw-kw+1)).astype('float32')
+    out_val = numpy.random.random((b, f, ih - kh + 1,
+                                   iw - kw + 1)).astype('float32')
 
     conv = dnn.dnn_conv(img, kern)
     gw = theano.grad(conv.sum(), kern)
@@ -546,7 +548,8 @@ def test_dnn_conv_grad():
     kw = 2
     img_val = numpy.random.random((b, c, ih, iw)).astype('float32')
     kern_val = numpy.random.random((f, c, kh, kw)).astype('float32')
-    out_val = numpy.random.random((b, f, ih-kw+1, iw-kw+1)).astype('float32')
+    out_val = numpy.random.random((b, f, ih - kw + 1,
+                                   iw - kw + 1)).astype('float32')
 
     def dconv(img, kern, out):
         desc = dnn.GpuDnnConvDesc(border_mode='valid', subsample=(1, 1),

theano/tests/test_flake8.py

@@ -220,7 +220,6 @@ whitelist_flake8 = [
     "sandbox/cuda/tests/test_tensor_op.py",
     "sandbox/cuda/tests/test_extra_ops.py",
     "sandbox/cuda/tests/test_gemmcorr3d.py",
-    "sandbox/cuda/tests/test_dnn.py",
     "sandbox/cuda/tests/test_viewop.py",
     "sandbox/scan_module/scan_utils.py",
     "sandbox/scan_module/scan.py",