未验证 提交 1a4a90a8 编写于 作者: K Kaipeng Deng 提交者: GitHub

Merge pull request #16140 from tink2123/arc_function

Add the inverse trigonometric function
...@@ -293,6 +293,7 @@ paddle.fluid.layers.sigmoid (ArgSpec(args=['x', 'name'], varargs=None, keywords= ...@@ -293,6 +293,7 @@ paddle.fluid.layers.sigmoid (ArgSpec(args=['x', 'name'], varargs=None, keywords=
paddle.fluid.layers.logsigmoid (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '81ccb7acafd06c7728e11581f5d342e3')) paddle.fluid.layers.logsigmoid (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '81ccb7acafd06c7728e11581f5d342e3'))
paddle.fluid.layers.exp (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e6b3e769413d96aab4176f96db25984b')) paddle.fluid.layers.exp (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e6b3e769413d96aab4176f96db25984b'))
paddle.fluid.layers.tanh (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e9d586a0b5bd05f67ee78048f9d503b6')) paddle.fluid.layers.tanh (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e9d586a0b5bd05f67ee78048f9d503b6'))
paddle.fluid.layers.atan (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '3a46e0b5f9ce82348406478e610f14c9'))
paddle.fluid.layers.tanh_shrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '1e521554b9fdda9061ec6d306f0709b7')) paddle.fluid.layers.tanh_shrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '1e521554b9fdda9061ec6d306f0709b7'))
paddle.fluid.layers.softshrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '9eef31597bbafa2bd49691e072296e13')) paddle.fluid.layers.softshrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '9eef31597bbafa2bd49691e072296e13'))
paddle.fluid.layers.sqrt (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '072a8541e0f632366bba10f67cb0db27')) paddle.fluid.layers.sqrt (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '072a8541e0f632366bba10f67cb0db27'))
...@@ -300,6 +301,8 @@ paddle.fluid.layers.abs (ArgSpec(args=['x', 'name'], varargs=None, keywords=None ...@@ -300,6 +301,8 @@ paddle.fluid.layers.abs (ArgSpec(args=['x', 'name'], varargs=None, keywords=None
paddle.fluid.layers.ceil (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c75d67dc5fe28f68e4cfffead4f698ad')) paddle.fluid.layers.ceil (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c75d67dc5fe28f68e4cfffead4f698ad'))
paddle.fluid.layers.floor (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '647b16c5da5ef909649ae02abb434973')) paddle.fluid.layers.floor (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '647b16c5da5ef909649ae02abb434973'))
paddle.fluid.layers.cos (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '485f2686bcc2fe37a4bd893769c8a3e2')) paddle.fluid.layers.cos (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '485f2686bcc2fe37a4bd893769c8a3e2'))
paddle.fluid.layers.acos (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '920a47734482276c069ba24c61c26b25'))
paddle.fluid.layers.asin (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'cf4ee2c9b9d7293556f8c5173dfb5d2c'))
paddle.fluid.layers.sin (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '01f1766aa76eff1df30147505b59f7c4')) paddle.fluid.layers.sin (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '01f1766aa76eff1df30147505b59f7c4'))
paddle.fluid.layers.round (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'b47f5da13913d3e56bdb1e612a73f3f2')) paddle.fluid.layers.round (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'b47f5da13913d3e56bdb1e612a73f3f2'))
paddle.fluid.layers.reciprocal (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'cc6ac2f14f03c52aaa83a59bf83b8d26')) paddle.fluid.layers.reciprocal (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'cc6ac2f14f03c52aaa83a59bf83b8d26'))
......
...@@ -13,7 +13,9 @@ See the License for the specific language governing permissions and ...@@ -13,7 +13,9 @@ See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#include "paddle/fluid/operators/activation_op.h" #include "paddle/fluid/operators/activation_op.h"
#include <memory>
#include <string> #include <string>
#include <unordered_map>
#include "paddle/fluid/operators/mkldnn/mkldnn_activation_op.h" #include "paddle/fluid/operators/mkldnn/mkldnn_activation_op.h"
#include "paddle/fluid/platform/port.h" #include "paddle/fluid/platform/port.h"
#ifdef PADDLE_WITH_CUDA #ifdef PADDLE_WITH_CUDA
...@@ -269,6 +271,48 @@ $$out = \\frac{x}{1 + \|x\|}$$ ...@@ -269,6 +271,48 @@ $$out = \\frac{x}{1 + \|x\|}$$
)DOC"; )DOC";
class AcosOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X", "Input of acos operator");
AddOutput("Out", "Output of acos operator");
AddComment(R"DOC(
Arccosine Activation Operator.
$$out = \cos^{-1}(x)$$
)DOC");
}
};
class AsinOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X", "Input of asin operator");
AddOutput("Out", "Output of asin operator");
AddComment(R"DOC(
Arcsine Activation Operator.
$$out = \sin^{-1}(x)$$
)DOC");
}
};
class AtanOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X", "Input of atan operator");
AddOutput("Out", "Output of atan operator");
AddComment(R"DOC(
Arctanh Activation Operator.
$$out = \tanh^{-1}(x)$$
)DOC");
}
};
class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker { class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker {
public: public:
void Make() override { void Make() override {
...@@ -543,7 +587,10 @@ namespace ops = paddle::operators; ...@@ -543,7 +587,10 @@ namespace ops = paddle::operators;
__macro(SoftShrink, softshrink); \ __macro(SoftShrink, softshrink); \
__macro(Abs, abs); \ __macro(Abs, abs); \
__macro(Cos, cos); \ __macro(Cos, cos); \
__macro(Acos, acos); \
__macro(Sin, sin); \ __macro(Sin, sin); \
__macro(Asin, asin); \
__macro(Atan, atan); \
__macro(Round, round); \ __macro(Round, round); \
__macro(Log, log); \ __macro(Log, log); \
__macro(Square, square); \ __macro(Square, square); \
......
...@@ -39,9 +39,8 @@ namespace operators { ...@@ -39,9 +39,8 @@ namespace operators {
Please refer to the layer_helper.py and get the details. Please refer to the layer_helper.py and get the details.
*/ */
static std::unordered_set<std::string> InplaceOpSet = { static std::unordered_set<std::string> InplaceOpSet = {
"sigmoid", "exp", "relu", "tanh", "sqrt", "ceil", "sigmoid", "exp", "relu", "tanh", "sqrt", "ceil",
"floor", "reciprocal", "relu6", "soft_relu", "hard_sigmoid", "floor", "reciprocal", "relu6", "soft_relu", "hard_sigmoid"};
};
static bool IsInplace(const std::string& op) { static bool IsInplace(const std::string& op) {
bool inplace = InplaceOpSet.count(op); bool inplace = InplaceOpSet.count(op);
...@@ -553,6 +552,101 @@ struct SinFunctor : public BaseActivationFunctor<T> { ...@@ -553,6 +552,101 @@ struct SinFunctor : public BaseActivationFunctor<T> {
} }
}; };
template <typename T>
struct Acos {
HOSTDEVICE T operator()(const T& val) const { return acos(val); }
};
template <>
struct Acos<platform::float16> {
HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
return platform::float16(acos(static_cast<float>(val)));
}
};
// Acos(x) = acos(x)
template <typename T>
struct AcosFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out>
void operator()(Device d, X x, Out out) const {
out.device(d) = x.unaryExpr(Acos<T>());
}
};
// acos'(x) = -1/sqrt(1-x^2)
template <typename T>
struct AcosGradFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out, typename dOut,
typename dX>
void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
dx.device(d) =
-dout * static_cast<T>(1) / (static_cast<T>(1) - x.square()).sqrt();
}
};
template <typename T>
struct Asin {
HOSTDEVICE T operator()(const T& val) const { return asin(val); }
};
template <>
struct Asin<platform::float16> {
HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
return platform::float16(asin(static_cast<float>(val)));
}
};
// Asin(x) = asin(x)
template <typename T>
struct AsinFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out>
void operator()(Device d, X x, Out out) const {
out.device(d) = x.unaryExpr(Asin<T>());
}
};
// asin'(x) = 1/sqrt(1-x^2)
template <typename T>
struct AsinGradFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out, typename dOut,
typename dX>
void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
dx.device(d) =
dout * static_cast<T>(1) / (static_cast<T>(1) - x.square()).sqrt();
}
};
template <typename T>
struct Atan {
HOSTDEVICE T operator()(const T& val) const { return atan(val); }
};
template <>
struct Atan<platform::float16> {
HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
return platform::float16(atan(static_cast<float>(val)));
}
};
// Atan(x) = atan(x)
template <typename T>
struct AtanFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out>
void operator()(Device d, X x, Out out) const {
out.device(d) = x.unaryExpr(Atan<T>());
}
};
// atan'(x) = 1 / (1 + x^2)
template <typename T>
struct AtanGradFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out, typename dOut,
typename dX>
void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
dx.device(d) = dout * static_cast<T>(1) / (static_cast<T>(1) + x.square());
}
};
// round(x) = [x] // round(x) = [x]
template <typename T> template <typename T>
struct RoundFunctor : public BaseActivationFunctor<T> { struct RoundFunctor : public BaseActivationFunctor<T> {
...@@ -1001,13 +1095,16 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> { ...@@ -1001,13 +1095,16 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> {
__macro(relu, ReluFunctor, ReluGradFunctor); \ __macro(relu, ReluFunctor, ReluGradFunctor); \
__macro(gelu, GeluFunctor, GeluGradFunctor); \ __macro(gelu, GeluFunctor, GeluGradFunctor); \
__macro(tanh, TanhFunctor, TanhGradFunctor); \ __macro(tanh, TanhFunctor, TanhGradFunctor); \
__macro(atan, AtanFunctor, AtanGradFunctor); \
__macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor); \ __macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor); \
__macro(sqrt, SqrtFunctor, SqrtGradFunctor); \ __macro(sqrt, SqrtFunctor, SqrtGradFunctor); \
__macro(abs, AbsFunctor, AbsGradFunctor); \ __macro(abs, AbsFunctor, AbsGradFunctor); \
__macro(ceil, CeilFunctor, ZeroGradFunctor); \ __macro(ceil, CeilFunctor, ZeroGradFunctor); \
__macro(floor, FloorFunctor, ZeroGradFunctor); \ __macro(floor, FloorFunctor, ZeroGradFunctor); \
__macro(cos, CosFunctor, CosGradFunctor); \ __macro(cos, CosFunctor, CosGradFunctor); \
__macro(acos, AcosFunctor, AcosGradFunctor); \
__macro(sin, SinFunctor, SinGradFunctor); \ __macro(sin, SinFunctor, SinGradFunctor); \
__macro(asin, AsinFunctor, AsinGradFunctor); \
__macro(round, RoundFunctor, ZeroGradFunctor); \ __macro(round, RoundFunctor, ZeroGradFunctor); \
__macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor); \ __macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor); \
__macro(log, LogFunctor, LogGradFunctor); \ __macro(log, LogFunctor, LogGradFunctor); \
......
...@@ -23,6 +23,7 @@ __activations_noattr__ = [ ...@@ -23,6 +23,7 @@ __activations_noattr__ = [
'logsigmoid', 'logsigmoid',
'exp', 'exp',
'tanh', 'tanh',
'atan',
'tanh_shrink', 'tanh_shrink',
'softshrink', 'softshrink',
'sqrt', 'sqrt',
...@@ -30,6 +31,8 @@ __activations_noattr__ = [ ...@@ -30,6 +31,8 @@ __activations_noattr__ = [
'ceil', 'ceil',
'floor', 'floor',
'cos', 'cos',
'acos',
'asin',
'sin', 'sin',
'round', 'round',
'reciprocal', 'reciprocal',
......
...@@ -100,6 +100,23 @@ class TestTanh(TestActivation): ...@@ -100,6 +100,23 @@ class TestTanh(TestActivation):
self.check_grad(['X'], 'Out', max_relative_error=0.007) self.check_grad(['X'], 'Out', max_relative_error=0.007)
class TestAtan(TestActivation):
def setUp(self):
self.op_type = "atan"
self.init_dtype()
x = np.random.uniform(0.1, 1, [11, 17]).astype(self.dtype)
out = np.arctan(x)
self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
self.outputs = {'Out': out}
def test_check_grad(self):
if self.dtype == np.float16:
return
self.check_grad(['X'], 'Out', max_relative_error=0.007)
class TestTanhShrink(TestActivation): class TestTanhShrink(TestActivation):
def setUp(self): def setUp(self):
self.op_type = "tanh_shrink" self.op_type = "tanh_shrink"
...@@ -248,6 +265,23 @@ class TestCos(TestActivation): ...@@ -248,6 +265,23 @@ class TestCos(TestActivation):
self.check_grad(['X'], 'Out', max_relative_error=0.007) self.check_grad(['X'], 'Out', max_relative_error=0.007)
class TestAcos(TestActivation):
def setUp(self):
self.op_type = "acos"
self.init_dtype()
x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype)
out = np.arccos(x)
self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
self.outputs = {'Out': out}
def test_check_grad(self):
if self.dtype == np.float16:
return
self.check_grad(['X'], 'Out', max_relative_error=0.007)
class TestSin(TestActivation): class TestSin(TestActivation):
def setUp(self): def setUp(self):
self.op_type = "sin" self.op_type = "sin"
...@@ -265,6 +299,23 @@ class TestSin(TestActivation): ...@@ -265,6 +299,23 @@ class TestSin(TestActivation):
self.check_grad(['X'], 'Out', max_relative_error=0.007) self.check_grad(['X'], 'Out', max_relative_error=0.007)
class TestAsin(TestActivation):
def setUp(self):
self.op_type = "asin"
self.init_dtype()
x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype)
out = np.arcsin(x)
self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
self.outputs = {'Out': out}
def test_check_grad(self):
if self.dtype == np.float16:
return
self.check_grad(['X'], 'Out', max_relative_error=0.007)
class TestRound(TestActivation): class TestRound(TestActivation):
def setUp(self): def setUp(self):
self.op_type = "round" self.op_type = "round"
...@@ -665,7 +716,10 @@ create_test_act_fp16_class(TestAbs) ...@@ -665,7 +716,10 @@ create_test_act_fp16_class(TestAbs)
create_test_act_fp16_class(TestCeil, grad_check=False) create_test_act_fp16_class(TestCeil, grad_check=False)
create_test_act_fp16_class(TestFloor, grad_check=False) create_test_act_fp16_class(TestFloor, grad_check=False)
create_test_act_fp16_class(TestCos, grad_atol=0.85) create_test_act_fp16_class(TestCos, grad_atol=0.85)
create_test_act_fp16_class(TestAcos, grad_atol=0.85)
create_test_act_fp16_class(TestSin) create_test_act_fp16_class(TestSin)
create_test_act_fp16_class(TestAsin)
create_test_act_fp16_class(TestAtan)
create_test_act_fp16_class(TestRound, grad_check=False) create_test_act_fp16_class(TestRound, grad_check=False)
create_test_act_fp16_class(TestRelu) create_test_act_fp16_class(TestRelu)
create_test_act_fp16_class(TestGelu) create_test_act_fp16_class(TestGelu)
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册