Python cuda.gpu_from_host函数代码示例

def test_reject_rect():
    for cls in (FilterActs, ImageActs):
        # Tests that running FilterActs with a non-square
        # kernel is an error
        rng = np.random.RandomState([2012, 10, 9])
        batch_size = 5
        rows = 10
        cols = 9
        channels = 3
        filter_rows = 4
        filter_cols = filter_rows + 1
        num_filters = 6

        images = shared(rng.uniform(-1., 1., (channels, rows, cols,
            batch_size)).astype('float32'), name='images')
        filters = shared(rng.uniform(-1., 1., (channels, filter_rows,
            filter_cols, num_filters)).astype('float32'), name='filters')

        gpu_images = gpu_from_host(images)
        gpu_filters = gpu_from_host(filters)

        if cls is ImageActs:
            output = cls()(gpu_images, gpu_filters,
                        as_tensor_variable((rows, cols)))
        else:
            output = cls()(gpu_images, gpu_filters)

        f = function([], output)
        try:
            output = f()
        except ValueError:
            continue
        assert False

 def local_to_gpu(node):
     """
     op(host_from_gpu()) -> host_from_gpu(op)
     gpu_from_host(op) -> op(gpu_from_host)
     """
     if isinstance(node.op, op):
         #op(host_from_gpu()) -> host_from_gpu(op)
         #If any of the input that go on the GPU are on the GPU,
         #move the op to the gpu.
         if any(node.inputs[idx].owner and
                isinstance(node.inputs[idx].owner.op, cuda.HostFromGpu)
                for idx in to_gpu):
             new_inp = list(node.inputs)
             for idx in to_gpu:
                 new_inp[idx] = cuda.gpu_from_host(new_inp[idx])
             return [cuda.host_from_gpu(op()(*new_inp))]
     if node.op == cuda.gpu_from_host:
         #gpu_from_host(op) -> op(gpu_from_host)
         host_input = node.inputs[0]
         if host_input.owner and isinstance(host_input.owner.op,
                                            op):
             op_node = host_input.owner
             new_inp = list(op_node.inputs)
             for idx in to_gpu:
                 new_inp[idx] = cuda.gpu_from_host(new_inp[idx])
             return [op()(*new_inp)]
     return False

    def local_to_gpu(node):
        """
        op(host_from_gpu()) -> host_from_gpu(op)
        gpu_from_host(op) -> op(gpu_from_host)

        """
        if isinstance(node.op, op):
            # op(host_from_gpu()) -> host_from_gpu(op)
            # If any of the input that go on the GPU are on the GPU,
            # move the op to the gpu.
            if any(node.inputs[idx].owner and
                   isinstance(node.inputs[idx].owner.op, cuda.HostFromGpu)
                   for idx in to_gpu):
                new_inp = list(node.inputs)
                for idx in to_gpu:
                    new_inp[idx] = cuda.gpu_from_host(new_inp[idx])
                result_node = op()(*new_inp)
                copy_stack_trace(node.outputs[0], result_node)
                transfer_node = result_node.transfer('cpu')
                copy_stack_trace(node.outputs[0], transfer_node)
                return [transfer_node]
        if node.op == cuda.gpu_from_host:
            # gpu_from_host(op) -> op(gpu_from_host)
            host_input = node.inputs[0]
            if host_input.owner and isinstance(host_input.owner.op,
                                               op):
                op_node = host_input.owner
                new_inp = list(op_node.inputs)
                for idx in to_gpu:
                    new_inp[idx] = cuda.gpu_from_host(new_inp[idx])
                new_node = op()(*new_inp)
                copy_stack_trace(host_input, new_node)
                return [new_node]
        return False

def test_gpu_opt():
    if not cuda.cuda_available:
        # Skip test if cuda_ndarray is not available.
        from nose.plugins.skip import SkipTest
        raise SkipTest('Optional package cuda not available')

    # We test the case where we put the op on the gpu when the output
    # is moved to the gpu.
    p = tensor.fmatrix()
    u = tensor.fvector()
    m = multinomial.MultinomialFromUniform('auto')(p, u)
    assert m.dtype == 'float32', m.dtype
    m_gpu = cuda.gpu_from_host(m)

    f = function([p, u], m_gpu, allow_input_downcast=True, mode=get_mode(True))
    assert any([type(node.op) is multinomial.GpuMultinomialFromUniform
                for node in f.maker.fgraph.toposort()])
    pval = numpy.arange(10000 * 4, dtype='float32').reshape((10000, 4))+0.1
    pval = pval / pval.sum(axis=1)[:, None]
    uval = numpy.ones_like(pval[:, 0]) * 0.5
    mval = f(pval, uval)

    # Test with a row, it was failing in the past.
    r = tensor.frow()
    m = multinomial.MultinomialFromUniform('auto')(r, u)
    assert m.dtype == 'float32', m.dtype
    m_gpu = cuda.gpu_from_host(m)

    f = function([r, u], m_gpu, allow_input_downcast=True, mode=get_mode(True))
    assert any([type(node.op) is multinomial.GpuMultinomialFromUniform
                for node in f.maker.fgraph.toposort()])
    pval = numpy.arange(1 * 4, dtype='float32').reshape((1, 4))+0.1
    pval = pval / pval.sum(axis=1)[:, None]
    uval = numpy.ones_like(pval[:, 0]) * 0.5
    mval2 = f(pval, uval)

def test_alloc_memset_0():
    i = tensor.iscalar()
    z = numpy.zeros((1,), dtype='float32')
    o = numpy.ones((1,), dtype='float32')
    ones = numpy.ones((2,), dtype='float32')

    # Test with 0
    a = basic_ops.gpu_alloc(cuda.gpu_from_host(tensor.constant(z)), i)
    f = theano.function([i], a, mode=mode_with_gpu)
    topo = f.maker.fgraph.toposort()
    assert len(topo) == 1
    assert isinstance(topo[0].op, basic_ops.GpuAlloc) and topo[0].op.memset_0
    assert (numpy.asarray(f(6)) == 0).all()

    # Test with 1
    a = basic_ops.gpu_alloc(cuda.gpu_from_host(tensor.constant(o)), i)
    f = theano.function([i], a, mode=mode_with_gpu)
    topo = f.maker.fgraph.toposort()
    assert len(topo) == 1
    assert isinstance(topo[0].op, basic_ops.GpuAlloc)
    assert not topo[0].op.memset_0
    assert (numpy.asarray(f(6)) == 1).all()

    # Test with 1, 1
    a = basic_ops.gpu_alloc(cuda.gpu_from_host(tensor.constant(ones)), i)
    f = theano.function([i], a, mode=mode_with_gpu)
    topo = f.maker.fgraph.toposort()
    assert len(topo) == 1
    assert isinstance(topo[0].op, basic_ops.GpuAlloc)
    assert not topo[0].op.memset_0
    assert (numpy.asarray(f(2)) == 1).all()

def test_reject_bad_filt_number():
    for cls in (FilterActs, ImageActs):
        # Tests that running FilterActs with a # of filters per
        # group that is not 16 is an error
        rng = np.random.RandomState([2012, 10, 9])
        batch_size = 5
        rows = 10
        cols = 9
        channels = 3
        filter_rows = 4
        filter_cols = filter_rows
        num_filters = 6

        images = shared(rng.uniform(-1., 1., (channels, rows, cols,
            batch_size)).astype('float32'), name='images')
        filters = shared(rng.uniform(-1., 1., (channels, filter_rows,
            filter_cols, num_filters)).astype('float32'), name='filters')

        gpu_images = gpu_from_host(images)
        gpu_filters = gpu_from_host(filters)

        output = cls()(gpu_images, gpu_filters)
        f = function([], output)
        try:
            output = f()
        except ValueError:
            continue
        assert False

def test_grad():

    rng = np.random.RandomState([2012, 10, 9])

    batch_size = 5
    rows = 10
    cols = 9
    channels = 3
    filter_rows = 4
    filter_cols = filter_rows
    num_filters = 16

    images = shared(rng.uniform(-1.0, 1.0, (channels, rows, cols, batch_size)).astype("float32"), name="images")
    filters = shared(
        rng.uniform(-1.0, 1.0, (channels, filter_rows, filter_cols, num_filters)).astype("float32"), name="filters"
    )

    gpu_images = gpu_from_host(images)
    gpu_filters = gpu_from_host(filters)

    output = FilterActs()(gpu_images, gpu_filters)
    output = host_from_gpu(output)
    # XXX: use verify_grad
    output_grad = grad(output.sum(), images)

    images_bc01 = images.dimshuffle(3, 0, 1, 2)
    filters_bc01 = filters.dimshuffle(3, 0, 1, 2)
    filters_bc01 = filters_bc01[:, :, ::-1, ::-1]

    output_conv2d = conv2d(images_bc01, filters_bc01, border_mode="valid")

    output_conv2d = output_conv2d.dimshuffle(1, 2, 3, 0)
    # XXX: use verify_grad
    output_conv2d_grad = grad(output_conv2d.sum(), images)
    f = function([], [output_grad, output_conv2d_grad])

    output_grad, output_conv2d_grad = f()

    warnings.warn(
        """test_match_valid_conv success criterion is not very strict. Can we verify that this is OK?
                     One possibility is that theano is numerically unstable and Alex's code is better.
                     Probably theano CPU 64 bit is OK but it's worth checking the others."""
    )
    if np.abs(output_grad - output_conv2d_grad).max() > 7.7e-6:
        assert type(output_grad) == type(output_conv2d_grad)
        assert output_grad.dtype == output_conv2d_grad.dtype
        if output_grad.shape != output_conv2d_grad.shape:
            print "cuda-convnet shape: ", output_grad.shape
            print "theano shape: ", output_conv2d_grad.shape
            assert False
        err = np.abs(output_grad - output_conv2d_grad)
        print "absolute error range: ", (err.min(), err.max())
        print "mean absolute error: ", err.mean()
        print "cuda-convnet value range: ", (output_grad.min(), output_grad.max())
        print "theano value range: ", (output_conv2d_grad.min(), output_conv2d_grad.max())
        assert False

def test_match_valid_conv_strided():

    # Tests that running FilterActs with stride is the same as running
    # theano's conv2D in valid mode and then downsampling

    rng = np.random.RandomState([2012,10,9])

    batch_size = 5
    rows = 9
    cols = 9
    channels = 3
    filter_rows = 3
    filter_cols = filter_rows
    stride = 3
    num_filters = 16

    images = shared(rng.uniform(-1., 1., (channels, rows, cols,
        batch_size)).astype('float32'), name='images')
    filters = shared(rng.uniform(-1., 1., (channels, filter_rows,
        filter_cols, num_filters)).astype('float32'), name='filters')

    gpu_images = gpu_from_host(images)
    gpu_filters = gpu_from_host(filters)

    output = FilterActs(stride=stride)(gpu_images, gpu_filters)
    output = host_from_gpu(output)

    images_bc01 = images.dimshuffle(3,0,1,2)
    filters_bc01 = filters.dimshuffle(3,0,1,2)
    filters_bc01 = filters_bc01[:,:,::-1,::-1]

    output_conv2d = conv2d(images_bc01, filters_bc01,
            border_mode='valid', subsample=(stride, stride))

    output_conv2d_orig = output_conv2d.dimshuffle(1,2,3,0)
    output_conv2d = output_conv2d_orig  # [:, ::stride, ::stride, :]
    f = function([], [output, output_conv2d, output_conv2d_orig])

    output, output_conv2d, output_conv2d_orig = f()

    warnings.warn("""test_match_valid_conv success criterion is not very strict. Can we verify that this is OK?
                     One possibility is that theano is numerically unstable and Alex's code is better.
                     Probably theano CPU 64 bit is OK but it's worth checking the others.""")
    if np.abs(output - output_conv2d).max() > 2.4e-6:
        assert type(output) == type(output_conv2d)
        assert output.dtype == output_conv2d.dtype
        if output.shape != output_conv2d.shape:
            print 'cuda-convnet shape: ',output.shape
            print 'theano shape: ',output_conv2d.shape
            assert False
        err = np.abs(output - output_conv2d)
        print 'absolute error range: ', (err.min(), err.max())
        print 'mean absolute error: ', err.mean()
        print 'cuda-convnet value range: ', (output.min(), output.max())
        print 'theano value range: ', (output_conv2d.min(), output_conv2d.max())
        assert False

def test_match_valid_conv():

    # Tests that running FilterActs with no padding is the same as running
    # theano's conv2D in valid mode

    rng = np.random.RandomState([2012,10,9])

    batch_size = 5
    rows = 10
    cols = 9
    channels = 3
    filter_rows = 4
    filter_cols = filter_rows
    num_filters = 16

    images = shared(rng.uniform(-1., 1., (channels, rows, cols,
        batch_size)).astype('float32'), name='images')
    filters = shared(rng.uniform(-1., 1., (channels, filter_rows,
        filter_cols, num_filters)).astype('float32'), name='filters')

    gpu_images = gpu_from_host(images)
    gpu_filters = gpu_from_host(filters)

    output = FilterActs()(gpu_images, gpu_filters)
    output = host_from_gpu(output)

    images_bc01 = images.dimshuffle(3,0,1,2)
    filters_bc01 = filters.dimshuffle(3,0,1,2)
    filters_bc01 = filters_bc01[:,:,::-1,::-1]

    output_conv2d = conv2d(images_bc01, filters_bc01,
            border_mode='valid')

    output_conv2d = output_conv2d.dimshuffle(1,2,3,0)

    try:
        f = function([], [output, output_conv2d])
    except:
        raise KnownFailureTest("cuda-convnet code depends on an unmerged theano feature.")

    output, output_conv2d = f()

    warnings.warn("test_match_valid_conv success criterion is not very strict. Can we verify that this is OK?")
    if np.abs(output - output_conv2d).max() > 2.4e-6:
        assert type(output) == type(output_conv2d)
        assert output.dtype == output_conv2d.dtype
        if output.shape != output_conv2d.shape:
            print 'cuda-convnet shape: ',output.shape
            print 'theano shape: ',output_conv2d.shape
            assert False
        err = np.abs(output - output_conv2d)
        print 'absolute error range: ', (err.min(), err.max())
        print 'mean absolute error: ', err.mean()
        print 'cuda-convnet value range: ', (output.min(), output.max())
        print 'theano value range: ', (output_conv2d.min(), output_conv2d.max())
        assert False

def insert_gpu_filter_acts(node):
    if isinstance(node.op, FilterActs):
        images, filters = node.inputs
        if any_from_gpu(images, filters) or any_gpu_client(*node.outputs):
            gpu_filter_acts = GpuFilterActs(
                    module_stride=node.op.module_stride,
                    partial_sum=1)
            return [host_from_gpu(gpu_filter_acts(
                gpu_from_host(images),
                gpu_from_host(filters)))]

def insert_gpu_weight_acts(node):
    if isinstance(node.op, WeightActs):
        images, hidacts, frows, fcols = node.inputs
        if any_from_gpu(images, hidacts) or any_gpu_client(*node.outputs):
            gpu_weight_acts = GpuWeightActs(
                    module_stride=node.op.module_stride,
                    partial_sum=1)
            return [host_from_gpu(gpu_weight_acts(
                gpu_from_host(images),
                gpu_from_host(hidacts),
                frows,
                fcols,
                ))]

def insert_gpu_img_acts(node):
    if isinstance(node.op, ImgActs):
        filters, hidacts, irows, icols = node.inputs
        if any_from_gpu(filters, hidacts) or any_gpu_client(*node.outputs):
            gpu_img_acts = GpuImgActs(
                    module_stride=node.op.module_stride,
                    partial_sum=1)
            return [host_from_gpu(gpu_img_acts(
                gpu_from_host(filters),
                gpu_from_host(hidacts),
                irows,
                icols,
                ))]

def traverse(out, x, x_copy, d, visited=None):
    ''' Function used by scan to parse the tree and figure out which nodes
    it needs to replace. There are two options :
        1) x and x_copy or on host, then you would replace x with x_copy
        2) x is on gpu, x_copy on host, then you need to replace
        host_from_gpu(x) with x_copy
    This happens because initially shared variables are on GPU .. which is
    fine for the main computational graph but confuses things a bit for the
    inner graph of scan '''
    # ``visited`` is a set of nodes that are already known and don't need to be
    # checked again, speeding up the traversal of multiply-connected graphs.
    # if a ``visited`` set is given, it will be updated in-place so the callee
    # knows which nodes we have seen.
    if visited is None:
        visited = set()
    if out in visited:
        return d
    visited.add(out)
    import theano.sandbox.cuda as cuda
    if out == x:
        d[out] = cuda.gpu_from_host(x_copy)
        return d
    elif out.owner is None:
        return d
    elif (cuda.cuda_available and
          out.owner.op == cuda.host_from_gpu and
          out.owner.inputs == [x]):
        d[out] = tensor.as_tensor_variable(x_copy)
        return d
    else:
        for inp in out.owner.inputs:
            d = traverse(inp, x, x_copy, d, visited)
        return d

    def lmul(self, x):
        """
        dot(x, A)
        aka, do convolution with input image x

        """

        check_cuda(str(type(self)) + ".lmul")
        # TODO Why is it CPU??
        print "Por que?!?!", type(x)
        cpu = "Cuda" not in str(type(x))
        if cpu:
            x = gpu_from_host(x)

        assert x.ndim == 5
        x_axes = self.input_axes
        assert len(x_axes) == 5

        op_axes = ("c", 0, 1, "t", "b")
        if tuple(x_axes) != op_axes:
            print "ssssssssssssssss"
            x = x.dimshuffle(*[x_axes.index(axis) for axis in op_axes])

        _x_4d_shape = (
            self.signal_shape[0],
            self.signal_shape[1],
            self.signal_shape[2],
            self.signal_shape[3] * self.signal_shape[4],
        )

        x = x.reshape(_x_4d_shape)

        x = gpu_contiguous(x)

        rval = FilterActs(self.pad, self.partial_sum, self.kernel_stride[0])(x, self._filters)

        if cpu:
            rval = host_from_gpu(rval)

        rval = rval.reshape(
            (
                self.filter_shape[3],
                self.filter_shape[4],
                rval.shape[1],
                rval.shape[2],
                self.signal_shape[3],
                self.signal_shape[4],
            )
        )

        rval = diagonal_subtensor(rval, 4, 0).sum(axis=0)

        # Format the output based on the output space
        rval_axes = self.output_axes
        assert len(rval_axes) == 5

        if tuple(rval_axes) != op_axes:
            rval = rval.dimshuffle(*[op_axes.index(axis) for axis in rval_axes])

        return rval

    def lmul(self, x):
        """
        dot(x, A)
        aka, do convolution with input image x

        """

        check_cuda(str(type(self)) + ".lmul")
        cpu = 'Cuda' not in str(type(x))
        #assert cpu
        if cpu:
            x = gpu_from_host(x)

        assert x.ndim == 5
        x_axes = self.input_axes
        assert len(x_axes) == 5

        #x = shapeprint(x)
        op_axes = ('b', 'c', 0, 1, 't')
        print x_axes, op_axes
        #if tuple(x_axes) != op_axes:
        #    x = x.dimshuffle(*[x_axes.index(axis) for axis in op_axes])

        #x = shapeprint(x)
        #self._filters = shapeprint(self._filters)

        rval = cuda.blas.GpuCorr3dMM(border_mode= 'valid',
                                     subsample = tuple(self.kernel_stride),
                                     pad=tuple(self.pad))(x, self._filters)
        #rval = conv3d(im, filt, None, None, (self.kernel_stride[0], self.kernel_stride[1]) )
        #rval = rval.dimshuffle(0,4,1,2,3)
        #print "hello"
        return rval