Skip to content

P
pytensor
提交 a0653671 authored 作者: Arjun Jain's avatar Arjun Jain
浏览文件
操作

Feature: support for all stride sized. This makes it work for all subsample (x,…

Feature: support for all stride sized. This makes it work for all subsample (x, y) values and all kernel, batch, image size compatible on GPU. @nouiz: Can you please run the tests once again just to be double sure? I think it works correctly.
上级 1b680cc3
隐藏空白字符变更
内嵌 并排
正在显示 包含 41 行增加 和 46 行删除
theano/sandbox/cuda/blas.py
浏览文件 @ a0653671
...@@ -517,9 +517,6 @@ class GpuCorrMM(GpuOp): ...@@ -517,9 +517,6 @@ class GpuCorrMM(GpuOp):
if pad != 0: if pad != 0:
raise NotImplementedError( raise NotImplementedError(
"GpuCorrMM don't implement the pad parameter") "GpuCorrMM don't implement the pad parameter")
if subsample != (1, 1):
raise NotImplementedError(
"GpuCorrMM we don't implement the subsample parameter")
def __eq__(self, other): def __eq__(self, other):
return type(self) == type(other) \ return type(self) == type(other) \
...@@ -662,7 +659,7 @@ class GpuCorrMM(GpuOp): ...@@ -662,7 +659,7 @@ class GpuCorrMM(GpuOp):
} }
out2 = corrMM(%(img)s, %(kern)s, %(out)s, padH, padW); out2 = corrMM(%(img)s, %(kern)s, %(out)s, dx, dy, padH, padW);
if (out2==NULL){ if (out2==NULL){
%(fail)s %(fail)s
} }
......
theano/sandbox/cuda/conv_gemm.cu
浏览文件 @ a0653671
...@@ -92,7 +92,10 @@ void im2col(const float* data_im, const int channels, ...@@ -92,7 +92,10 @@ void im2col(const float* data_im, const int channels,
CudaNdarray* corrMM(const CudaNdarray *input, CudaNdarray* corrMM(const CudaNdarray *input,
CudaNdarray *weight, CudaNdarray *weight,
CudaNdarray *output, CudaNdarray *output,
int padH, int padW = 0) int dH = 1,
int dW = 1,
int padH = 0,
int padW = 0)
{ {
cublasStatus_t status; cublasStatus_t status;
...@@ -107,9 +110,6 @@ CudaNdarray* corrMM(const CudaNdarray *input, ...@@ -107,9 +110,6 @@ CudaNdarray* corrMM(const CudaNdarray *input,
PyErr_SetString(PyExc_ValueError, "required weight of 4D"); PyErr_SetString(PyExc_ValueError, "required weight of 4D");
} }
// TODO: stride(dW, dH) and padding as function parameter
int dH = 1;
int dW = 1;
int kH = CudaNdarray_HOST_DIMS(weight)[2]; int kH = CudaNdarray_HOST_DIMS(weight)[2];
int kW = CudaNdarray_HOST_DIMS(weight)[3]; int kW = CudaNdarray_HOST_DIMS(weight)[3];
int nInputPlane = CudaNdarray_HOST_DIMS(input)[1]; int nInputPlane = CudaNdarray_HOST_DIMS(input)[1];
......
theano/sandbox/cuda/tests/test_conv_cuda_ndarray.py
浏览文件 @ a0653671
...@@ -644,10 +644,6 @@ def test_valid(): ...@@ -644,10 +644,6 @@ def test_valid():
# Add tests with strided inputs by still square images and filters. # Add tests with strided inputs by still square images and filters.
shapes += get_shapes2(scales_img=(2, 2), img_stride=(2, 2)) shapes += get_shapes2(scales_img=(2, 2), img_stride=(2, 2))
shapes += get_shapes2(scales_kern=(2, 2), kern_stride=(2, 2)) shapes += get_shapes2(scales_kern=(2, 2), kern_stride=(2, 2))
# Keep only tests with square images and filters even with inputs strides
shapes = [shp for shp in shapes if (
shp[0][2]/shp[3][0] == shp[0][3]/shp[3][1] and
shp[1][2]/shp[4][0] == shp[1][3]/shp[4][1])]
exec_conv(version, shapes, verbose, random, 'valid', exec_conv(version, shapes, verbose, random, 'valid',
print_=print_, ones=ones, rtol=1.1e-5, print_=print_, ones=ones, rtol=1.1e-5,
theano_mode=mode, cls=cuda.blas.GpuCorrMM) theano_mode=mode, cls=cuda.blas.GpuCorrMM)
...@@ -718,8 +714,6 @@ def test_full(): ...@@ -718,8 +714,6 @@ def test_full():
# Test the GpuCorrMM version # Test the GpuCorrMM version
mode = theano_mode.including("conv_gemm") mode = theano_mode.including("conv_gemm")
shapes = [shp for shp in shapes if shp[1][2] == shp[1][3]]
shapes = [shp for shp in shapes if shp[0][2] == shp[0][3]]
shapes = shapes[0:10] shapes = shapes[0:10]
exec_conv(version, shapes, verbose, random, 'full', exec_conv(version, shapes, verbose, random, 'full',
theano_mode=mode, cls=cuda.blas.GpuCorrMM) theano_mode=mode, cls=cuda.blas.GpuCorrMM)
...@@ -841,38 +835,42 @@ def test_gemm(): ...@@ -841,38 +835,42 @@ def test_gemm():
for rImg2 in range(5, 9): for rImg2 in range(5, 9):
for rFlt1 in range(2, 4): for rFlt1 in range(2, 4):
for rFlt2 in range(2, 4): for rFlt2 in range(2, 4):
ishape = (bs, ch, rImg1, rImg2) for subsx in range(1, 3):
kshape = (nf, ch, rFlt1, rFlt2) for subsy in range(1, 3):
print "ishape: ", ishape ishape = (bs, ch, rImg1, rImg2)
print "kshape: ", kshape kshape = (nf, ch, rFlt1, rFlt2)
subsample = (1, 1) print "ishape: ", ishape
print "kshape: ", kshape
npy_img = theano._asarray(numpy.random.rand(*ishape), dtype='float32') subsample = (subsx, subsy)
npy_kern = theano._asarray(numpy.random.rand(*kshape), dtype='float32') print "subsample: ", subsample
i = cuda_tensor4() npy_img = theano._asarray(numpy.random.rand(*ishape), dtype='float32')
k = cuda_tensor4() npy_kern = theano._asarray(numpy.random.rand(*kshape), dtype='float32')
t2 = None i = cuda_tensor4()
k = cuda_tensor4()
t0 = time.time()
cpuval = py_conv(npy_img, npy_kern, mode, subsample) t2 = None
t1 = time.time() t0 = time.time()
cpuval = py_conv(npy_img, npy_kern, mode, subsample)
op = theano.sandbox.cuda.blas.GpuCorrMM(border_mode=mode)(i, k)
f = theano.function([i, k], op, mode=theano_mode) t1 = time.time()
npy_kern = npy_kern[:,:,::-1,::-1] op = theano.sandbox.cuda.blas.GpuCorrMM(border_mode=mode, \
subsample=subsample)(i, k)
gpuval = f(npy_img, npy_kern) f = theano.function([i, k], op, mode=theano_mode)
t2 = time.time() npy_kern = npy_kern[:,:,::-1,::-1]
gpuval = numpy.asarray(gpuval) gpuval = f(npy_img, npy_kern)
rval = numpy.allclose(cpuval, gpuval, rtol=1e-4)
assert (rval == True) t2 = time.time()
print 'Test Passed'
gpuval = numpy.asarray(gpuval)
rval = numpy.allclose(cpuval, gpuval, rtol=1e-4)
assert (rval == True)
print 'Test Passed'
def benchmark(): def benchmark():
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论