Skip to content

P
pytensor
提交 bae4eb91 authored 作者: Aleksandar Botev's avatar Aleksandar Botev
浏览文件
操作

Added mode 'half' to Images2Neibs. Tests pass. #5938

上级 22eaec56
全部展开 显示空白字符变更
内嵌 并排
正在显示 包含 125 行增加 和 23 行删除
theano/gpuarray/neighbours.py
浏览文件 @ bae4eb91
...@@ -23,9 +23,9 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op): ...@@ -23,9 +23,9 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
""" """
def __init__(self, mode='valid'): def __init__(self, mode='valid'):
if mode not in ['valid', 'ignore_borders', 'wrap_centered']: if mode not in ['valid', 'ignore_borders', 'wrap_centered', 'half']:
raise NotImplementedError("Only the mode valid, ignore_borders" raise NotImplementedError("Only the mode valid, ignore_borders"
" and wrap_centered" ", wrap_centered and half"
" have been implemented for the op" " have been implemented for the op"
" GpuImages2Neibs") " GpuImages2Neibs")
self.mode = mode self.mode = mode
...@@ -85,8 +85,8 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op): ...@@ -85,8 +85,8 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
GLOBAL_MEM %(type_z)s * global_out, const ga_size offset_out GLOBAL_MEM %(type_z)s * global_out, const ga_size offset_out
) )
{ {
const ga_int wrap_centered_idx_shift_x = c/2; const ga_int wrap_centered_half_idx_shift_x = c/2;
const ga_int wrap_centered_idx_shift_y = d/2; const ga_int wrap_centered_half_idx_shift_y = d/2;
global_ten4 = (GLOBAL_MEM const %(type_ten4)s *)(((GLOBAL_MEM char *)global_ten4)+offset_ten4); global_ten4 = (GLOBAL_MEM const %(type_ten4)s *)(((GLOBAL_MEM char *)global_ten4)+offset_ten4);
global_out = (GLOBAL_MEM %(type_z)s *)(((GLOBAL_MEM char *)global_out)+offset_out); global_out = (GLOBAL_MEM %(type_z)s *)(((GLOBAL_MEM char *)global_out)+offset_out);
...@@ -111,34 +111,41 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op): ...@@ -111,34 +111,41 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
{ {
ga_int ten4_2 = i + a * step_x; ga_int ten4_2 = i + a * step_x;
if("%(mode)s"=="wrap_centered"){ if("%(mode)s"=="wrap_centered"){
ten4_2 -= wrap_centered_idx_shift_x; ten4_2 -= wrap_centered_half_idx_shift_x;
if ( ten4_2 < 0 ) if ( ten4_2 < 0 )
ten4_2 += height; ten4_2 += height;
else if (ten4_2 >= height) else if (ten4_2 >= height)
ten4_2 -= height; ten4_2 -= height;
} else if ("%(mode)s"=="half"){
ten4_2 -= wrap_centered_half_idx_shift_x;
} }
ga_int j = LID_0; // loop over d ga_int j = LID_0; // loop over d
{ {
ga_int ten4_3 = j + b * step_y; ga_int ten4_3 = j + b * step_y;
if("%(mode)s"=="wrap_centered"){ if("%(mode)s"=="wrap_centered"){
ten4_3 -= wrap_centered_idx_shift_y; ten4_3 -= wrap_centered_half_idx_shift_y;
if ( ten4_3 < 0 ) if ( ten4_3 < 0 )
ten4_3 += width; ten4_3 += width;
else if (ten4_3 >= width) else if (ten4_3 >= width)
ten4_3 -= width; ten4_3 -= width;
} else if ("%(mode)s"=="half"){
ten4_3 -= wrap_centered_half_idx_shift_y;
} }
ga_int ten4_idx = stride3*ten4_3 +
stride2*ten4_2 +
stride1*s + stride0*n;
ga_int z_col = j + d * i; ga_int z_col = j + d * i;
ga_int z_idx = z_col * out_s1 + ga_int z_idx = z_col * out_s1 +
z_row * out_s0; z_row * out_s0;
if(ten4_2 < 0 || ten4_2 >= height || ten4_3 < 0 || ten4_3 >= width){
global_ot[z_idx] = 0;
} else {
ga_int ten4_idx = stride3*ten4_3 +
stride2*ten4_2 +
stride1*s + stride0*n;
global_out[z_idx] = global_ten4[ten4_idx]; global_out[z_idx] = global_ten4[ten4_idx];
} }
} }
} }
}
}""" % locals() }""" % locals()
params = [ params = [
'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc',
...@@ -172,8 +179,8 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op): ...@@ -172,8 +179,8 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
GLOBAL_MEM %(type_z)s * global_out, const ga_size offset_out GLOBAL_MEM %(type_z)s * global_out, const ga_size offset_out
) )
{ {
const ga_int wrap_centered_idx_shift_x = c/2; const ga_int wrap_centered_half_idx_shift_x = c/2;
const ga_int wrap_centered_idx_shift_y = d/2; const ga_int wrap_centered_half_idx_shift_y = d/2;
global_ten4 = (GLOBAL_MEM const %(type_ten4)s *)(((GLOBAL_MEM char *)global_ten4)+offset_ten4); global_ten4 = (GLOBAL_MEM const %(type_ten4)s *)(((GLOBAL_MEM char *)global_ten4)+offset_ten4);
global_out = (GLOBAL_MEM %(type_z)s *)(((GLOBAL_MEM char *)global_out)+offset_out); global_out = (GLOBAL_MEM %(type_z)s *)(((GLOBAL_MEM char *)global_out)+offset_out);
...@@ -199,36 +206,43 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op): ...@@ -199,36 +206,43 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
{ {
ga_int ten4_2 = i + a * step_x; ga_int ten4_2 = i + a * step_x;
if("%(mode)s"=="wrap_centered"){ if("%(mode)s"=="wrap_centered"){
ten4_2 -= wrap_centered_idx_shift_x; ten4_2 -= wrap_centered_half_idx_shift_x;
if ( ten4_2 < 0 ) if ( ten4_2 < 0 )
ten4_2 += height; ten4_2 += height;
else if (ten4_2 >= height) else if (ten4_2 >= height)
ten4_2 -= height; ten4_2 -= height;
} else if ("%(mode)s"=="half"){
ten4_2 -= wrap_centered_half_idx_shift_x;
} }
// loop over d // loop over d
for (ga_int j = LID_0; j < d; j+=LDIM_0) for (ga_int j = LID_0; j < d; j+=LDIM_0)
{ {
ga_int ten4_3 = j + b * step_y; ga_int ten4_3 = j + b * step_y;
if("%(mode)s"=="wrap_centered"){ if("%(mode)s"=="wrap_centered"){
ten4_3 -= wrap_centered_idx_shift_y; ten4_3 -= wrap_centered_half_idx_shift_y;
if ( ten4_3 < 0 ) if ( ten4_3 < 0 )
ten4_3 += width; ten4_3 += width;
else if (ten4_3 >= width) else if (ten4_3 >= width)
ten4_3 -= width; ten4_3 -= width;
} else if ("%(mode)s"=="half"){
ten4_3 -= wrap_centered_half_idx_shift_y;
} }
ga_int ten4_idx = stride3*ten4_3 +
stride2*ten4_2 +
stride1*s + stride0*n;
ga_int z_col = j + d * i; ga_int z_col = j + d * i;
ga_int z_idx = z_col * out_s1 + ga_int z_idx = z_col * out_s1 +
z_row * out_s0; z_row * out_s0;
if(ten4_2 < 0 || ten4_2 >= height || ten4_3 < 0 || ten4_3 >= width){
global_ot[z_idx] = 0;
} else {
ga_int ten4_idx = stride3*ten4_3 +
stride2*ten4_2 +
stride1*s + stride0*n;
global_out[z_idx] = global_ten4[ten4_idx]; global_out[z_idx] = global_ten4[ten4_idx];
} }
} }
} }
} }
}
""" % locals() """ % locals()
params = [ params = [
'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc', 'intc',
...@@ -367,6 +381,31 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op): ...@@ -367,6 +381,31 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
grid_c = 1+(((PyGpuArray_DIMS(%(ten4)s))[2]-c)/step_x); grid_c = 1+(((PyGpuArray_DIMS(%(ten4)s))[2]-c)/step_x);
//number of patch in width //number of patch in width
grid_d = 1+(((PyGpuArray_DIMS(%(ten4)s))[3]-d)/step_y); grid_d = 1+(((PyGpuArray_DIMS(%(ten4)s))[3]-d)/step_y);
}else if ( "%(mode)s" == "half") {
if ( ((PyGpuArray_DIMS(%(ten4)s))[2] < c) ||
((((PyGpuArray_DIMS(%(ten4)s))[2]-(c%%2)) %% step_x)!=0))
{
PyErr_Format(PyExc_TypeError, "GpuImages2Neibs:"
" neib_shape[0]=%%d, neib_step[0]=%%d and"
" ten4.shape[2]=%%d not consistent",
c, step_x,
PyGpuArray_DIMS(%(ten4)s)[2]);
%(fail)s;
}
if ( ((PyGpuArray_DIMS(%(ten4)s))[3] < d) ||
((((PyGpuArray_DIMS(%(ten4)s))[3]-(d%%2)) %% step_y)!=0))
{
PyErr_Format(PyExc_TypeError, "GpuImages2Neibs:"
" neib_shape[1]=%%d, neib_step[1]=%%d and"
" ten4.shape[3]=%%d not consistent",
d, step_y,
PyGpuArray_DIMS(%(ten4)s)[3]);
%(fail)s;
}
//number of patch in height
grid_c = 1+(((PyGpuArray_DIMS(%(ten4)s))[2]-(c%%2))/step_x);
//number of patch in width
grid_d = 1+(((PyGpuArray_DIMS(%(ten4)s))[3]-(d%%2))/step_y);
}else{ }else{
PyErr_Format(PyExc_TypeError, PyErr_Format(PyExc_TypeError,
"GpuImages2Neibs:: unknown mode '%(mode)s'"); "GpuImages2Neibs:: unknown mode '%(mode)s'");
...@@ -485,5 +524,5 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op): ...@@ -485,5 +524,5 @@ class GpuImages2Neibs(GpuKernelBase, Images2Neibs, Op):
@op_lifter([Images2Neibs]) @op_lifter([Images2Neibs])
@register_opt2([Images2Neibs], 'fast_compile') @register_opt2([Images2Neibs], 'fast_compile')
def local_gpua_images2neibs(op, context_name, inputs, outputs): def local_gpua_images2neibs(op, context_name, inputs, outputs):
if op.mode in ['valid', 'ignore_borders', 'wrap_centered']: if op.mode in ['valid', 'ignore_borders', 'wrap_centered', 'half']:
return GpuImages2Neibs(op.mode) return GpuImages2Neibs(op.mode)
theano/tensor/nnet/tests/test_neighbours.py
浏览文件 @ bae4eb91
...@@ -236,6 +236,29 @@ class T_Images2Neibs(unittest_tools.InferShapeTester): ...@@ -236,6 +236,29 @@ class T_Images2Neibs(unittest_tools.InferShapeTester):
# TODO: why this is commented? # TODO: why this is commented?
# assert numpy.allclose(images.get_value(borrow=True), g()) # assert numpy.allclose(images.get_value(borrow=True), g())
def test_neibs_half_step_by_valid(self):
for shp_idx, (shape, neib_shape, neib_step) in enumerate([
[(7, 8, 5, 5), (3, 3), (1, 1)],
[(7, 8, 5, 5), (3, 3), (2, 2)],
[(7, 8, 5, 5), (3, 3), (4, 4)],
[(7, 8, 5, 5), (3, 3), (1, 4)],
[(7, 8, 5, 5), (3, 3), (4, 1)],
[(80, 90, 5, 5), (3, 3), (1, 2)],
[(1025, 9, 5, 5), (3, 3), (2, 1)],
[(1, 1, 5, 1037), (3, 3), (2, 4)],
[(1, 1, 1045, 5), (3, 3), (4, 2)]]
):
for dtype in self.dtypes:
x = theano.shared(np.random.randn(*shape).astype(dtype))
extra = (neib_shape[0] // 2, neib_shape[1] // 2)
padded_shape = (x.shape[0], x.shape[1], x.shape[2] + 2 * extra[0], x.shape[3] + 2 * extra[1])
padded_x = T.zeros(padded_shape)
padded_x = T.set_subtensor(padded_x[:, :, extra[0]:-extra[0], extra[1]:-extra[1]], x)
x_using_valid = images2neibs(padded_x, neib_shape, neib_step, mode="valid")
x_using_half = images2neibs(x, neib_shape, neib_step, mode="half")
close = T.allclose(x_using_valid, x_using_half)
assert close.eval()
def test_neibs_bad_shape_wrap_centered(self): def test_neibs_bad_shape_wrap_centered(self):
shape = (2, 3, 10, 10) shape = (2, 3, 10, 10)
...@@ -281,6 +304,17 @@ class T_Images2Neibs(unittest_tools.InferShapeTester): ...@@ -281,6 +304,17 @@ class T_Images2Neibs(unittest_tools.InferShapeTester):
self.assertRaises(TypeError, unittest_tools.verify_grad, self.assertRaises(TypeError, unittest_tools.verify_grad,
fn, [images_val], mode=self.mode) fn, [images_val], mode=self.mode)
def test_grad_half(self):
# It is not implemented for now. So test that we raise an error.
shape = (2, 3, 6, 6)
images_val = np.random.rand(*shape).astype('float32')
def fn(images):
return images2neibs(images, (3, 3), mode='half')
self.assertRaises(TypeError, unittest_tools.verify_grad,
fn, [images_val], mode=self.mode)
def test_grad_valid(self): def test_grad_valid(self):
shape = (2, 3, 6, 6) shape = (2, 3, 6, 6)
images_val = np.random.rand(*shape).astype('float32') images_val = np.random.rand(*shape).astype('float32')
...@@ -330,15 +364,22 @@ class T_Images2Neibs(unittest_tools.InferShapeTester): ...@@ -330,15 +364,22 @@ class T_Images2Neibs(unittest_tools.InferShapeTester):
images_val = np.arange(np.prod(shape), images_val = np.arange(np.prod(shape),
dtype='float32').reshape(shape) dtype='float32').reshape(shape)
def fn(images):
return T.sum(T.sqr(images2neibs(images, (2, 2), mode='valid')),
axis=[0, 1])
f = theano.function([images], f = theano.function([images],
T.sqr(images2neibs(images, (2, 2), mode='valid')), T.sqr(images2neibs(images, (2, 2), mode='valid')),
mode=self.mode) mode=self.mode)
self.assertRaises(TypeError, f, images_val) self.assertRaises(TypeError, f, images_val)
def test_neibs_half_with_inconsistent_borders(self):
shape = (2, 3, 5, 5)
images = T.dtensor4()
images_val = np.arange(np.prod(shape),
dtype='float32').reshape(shape)
f = theano.function([images],
T.sqr(images2neibs(images, (2, 2), mode='half')),
mode=self.mode)
self.assertRaises(TypeError, f, images_val)
def test_can_not_infer_nb_dim(self): def test_can_not_infer_nb_dim(self):
# Was reported in gh-5613. Test that we do not crash # Was reported in gh-5613. Test that we do not crash
# or that we crash in a few other case found while # or that we crash in a few other case found while
...@@ -392,6 +433,19 @@ class T_Images2Neibs(unittest_tools.InferShapeTester): ...@@ -392,6 +433,19 @@ class T_Images2Neibs(unittest_tools.InferShapeTester):
for i in range(1000): for i in range(1000):
f() f()
def speed_neibs_half(self):
shape = (100, 40, 18, 18)
images = shared(np.arange(np.prod(shape),
dtype='float32').reshape(shape))
neib_shape = T.as_tensor_variable((3, 3))
f = function([],
images2neibs(images, neib_shape, mode="half"),
mode=self.mode)
for i in range(1000):
f()
def test_infer_shape(self): def test_infer_shape(self):
shape = (100, 40, 6, 3) shape = (100, 40, 6, 3)
images = np.ones(shape).astype('float32') images = np.ones(shape).astype('float32')
...@@ -431,6 +485,15 @@ class T_Images2Neibs(unittest_tools.InferShapeTester): ...@@ -431,6 +485,15 @@ class T_Images2Neibs(unittest_tools.InferShapeTester):
[x], [images2neibs( [x], [images2neibs(
x, neib_shape=(3, 3), mode='wrap_centered')], x, neib_shape=(3, 3), mode='wrap_centered')],
[images], Images2Neibs) [images], Images2Neibs)
shape = (100, 40, 6, 4)
images = np.ones(shape).astype('float32')
x = T.ftensor4()
self._compile_and_check(
[x], [images2neibs(x, neib_shape=(2, 1), mode='half')],
[images], Images2Neibs)
self._compile_and_check(
[x], [images2neibs(x, neib_shape=(2, 3), mode='half')],
[images], Images2Neibs)
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()
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