Add digamma_op and unittest

paddle/fluid/operators/digamma_op.cc

@@ -0,0 +1,100 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/digamma_op.h"
+
+namespace paddle {
+namespace operators {
+
+class DigammaOp : public framework::OperatorWithKernel {
+ public:
+  DigammaOp(const std::string &type, const framework::VariableNameMap &inputs,
+            const framework::VariableNameMap &outputs,
+            const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "Digamma");
+    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "Digamma");
+
+    auto in_dims = ctx->GetInputDim("X");
+
+    ctx->SetOutputDim("Out", in_dims);
+    ctx->ShareLoD("X", "Out");
+  }
+};
+
+class DigammaOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X", "(Tensor), The input tensor of digamma operator.");
+    AddOutput("Out", "(Tensor), The output tensor of digamma operator.");
+    AddComment(R"DOC(
+Digamma Operator.
+
+This operator is used to perform elementwise digamma for input $X$.
+$$out = \Psi(x) = \frac{ \Gamma^{'}(x) }{ \Gamma(x) }$$
+
+)DOC");
+  }
+};
+
+class DigammaGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
+                   "Out@Grad", "DigammaGrad");
+    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "DigammaGrad");
+    OP_INOUT_CHECK(ctx->HasOutput(framework::GradVarName("X")), "Output",
+                   "X@Grad", "DigammaGrad");
+
+    auto dout_dims = ctx->GetInputDim(framework::GradVarName("Out"));
+    ctx->SetOutputDim(framework::GradVarName("X"), dout_dims);
+    ctx->ShareLoD(framework::GradVarName("Out"), framework::GradVarName("X"));
+  }
+};
+
+template <typename T>
+class DigammaGradOpMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+  void Apply(GradOpPtr<T> retv) const override {
+    retv->SetType("digamma_grad");
+    retv->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    retv->SetInput("X", this->Input("X"));
+    retv->SetAttrMap(this->Attrs());
+    retv->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OPERATOR(digamma, ops::DigammaOp, ops::DigammaOpMaker,
+                  ops::DigammaGradOpMaker<paddle::framework::OpDesc>,
+                  ops::DigammaGradOpMaker<paddle::imperative::OpBase>);
+REGISTER_OPERATOR(digamma_grad, ops::DigammaGradOp);
+
+REGISTER_OP_CPU_KERNEL(
+    digamma, ops::DigammaKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::DigammaKernel<paddle::platform::CPUDeviceContext, double>);
+
+REGISTER_OP_CPU_KERNEL(
+    digamma_grad,
+    ops::DigammaGradKernel<paddle::platform::CPUDeviceContext, float>,
+    ops::DigammaGradKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/digamma_op.cu

@@ -0,0 +1,26 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/digamma_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_CUDA_KERNEL(
+    digamma, ops::DigammaKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::DigammaKernel<paddle::platform::CUDADeviceContext, double>);
+
+REGISTER_OP_CUDA_KERNEL(
+    digamma_grad,
+    ops::DigammaGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::DigammaGradKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/operators/digamma_op.h

@@ -0,0 +1,99 @@
+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include <unsupported/Eigen/SpecialFunctions>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/platform/for_range.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+struct DigammaFunctor {
+  DigammaFunctor(const T* input, T* output, int64_t numel)
+      : input_(input), output_(output), numel_(numel) {}
+
+  HOSTDEVICE void operator()(int64_t idx) const {
+    output_[idx] = Eigen::numext::digamma(input_[idx]);
+  }
+
+ private:
+  const T* input_;
+  T* output_;
+  int64_t numel_;
+};
+
+template <typename T>
+struct DigammaGradFunctor {
+  DigammaGradFunctor(const T* dout, const T* x, T* output, int64_t numel)
+      : dout_(dout), x_(x), output_(output), numel_(numel) {}
+
+  HOSTDEVICE void operator()(int64_t idx) const {
+    output_[idx] = dout_[idx] * Eigen::numext::polygamma(T(1), x_[idx]);
+  }
+
+ private:
+  const T* dout_;
+  const T* x_;
+  T* output_;
+  int64_t numel_;
+};
+
+using Tensor = framework::Tensor;
+
+template <typename DeviceContext, typename T>
+class DigammaKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* x = context.Input<Tensor>("X");
+    Tensor* out = context.Output<Tensor>("Out");
+
+    auto numel = x->numel();
+    auto* x_data = x->data<T>();
+    auto* out_data = out->mutable_data<T>(context.GetPlace(),
+                                          size_t(x->numel() * sizeof(T)));
+
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
+    DigammaFunctor<T> functor(x_data, out_data, numel);
+    for_range(functor);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class DigammaGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
+    const Tensor* x = context.Input<Tensor>("X");
+    auto* d_x = context.Output<Tensor>(framework::GradVarName("X"));
+
+    auto numel = d_out->numel();
+    auto* dout_data = d_out->data<T>();
+    auto* x_data = x->data<T>();
+    auto* dx_data = d_x->mutable_data<T>(
+        context.GetPlace(), static_cast<size_t>(numel * sizeof(T)));
+
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    platform::ForRange<DeviceContext> for_range(dev_ctx, numel);
+    DigammaGradFunctor<T> functor(dout_data, x_data, dx_data, numel);
+    for_range(functor);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

python/paddle/__init__.py

@@ -204,6 +204,7 @@
 from .tensor.math import prod  # noqa: F401
 from .tensor.math import broadcast_shape  # noqa: F401
 from .tensor.math import conj  # noqa: F401
+from .tensor.math import digamma  # noqa: F401
 from .tensor.math import neg  # noqa: F401
 from .tensor.math import lgamma  # noqa: F401
 
@@ -487,5 +488,6 @@
            'log10',
            'concat',
            'check_shape',
+           'digamma',
            'standard_normal'
 ]

python/paddle/fluid/tests/unittests/test_digamma_op.py

@@ -0,0 +1,119 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import math
+import numpy as np
+from scipy.special import psi
+import paddle
+import paddle.fluid as fluid
+import paddle.static as static
+from op_test import OpTest
+
+
+class TestDigammaOp(OpTest):
+    def setUp(self):
+        # switch to static
+        paddle.enable_static()
+
+        self.op_type = 'digamma'
+        self.init_dtype_type()
+        shape = (5, 32)
+        data = np.random.random(shape).astype(self.dtype) + 1
+        self.inputs = {'X': data}
+        result = np.ones(shape).astype(self.dtype)
+        result = psi(data)
+        self.outputs = {'Out': result}
+
+    def init_dtype_type(self):
+        self.dtype = np.float64
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X'], 'Out')
+
+
+class TestDigammaOpFp32(TestDigammaOp):
+    def init_dtype_type(self):
+        self.dtype = np.float32
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X'], 'Out')
+
+
+class TestDigammaAPI(unittest.TestCase):
+    def setUp(self):
+        # switch to static
+        paddle.enable_static()
+        # prepare test attrs
+        self.dtypes = ["float32", "float64"]
+        self.places = [paddle.CPUPlace()]
+        if paddle.is_compiled_with_cuda():
+            self.places.append(paddle.CUDAPlace(0))
+        self._shape = [8, 3, 32, 32]
+
+    def test_in_static_mode(self):
+        def init_input_output(dtype):
+            input = np.random.random(self._shape).astype(dtype)
+            return {'x': input}, psi(input)
+
+        for dtype in self.dtypes:
+            input_dict, sc_res = init_input_output(dtype)
+            for place in self.places:
+                with static.program_guard(static.Program()):
+                    x = static.data(name="x", shape=self._shape, dtype=dtype)
+                    out = paddle.digamma(x)
+
+                    exe = static.Executor(place)
+                    out_value = exe.run(feed=input_dict, fetch_list=[out.name])
+                    self.assertEqual(
+                        np.allclose(
+                            out_value[0], sc_res, rtol=1e-5), True)
+
+    def test_in_dynamic_mode(self):
+        for dtype in self.dtypes:
+            input = np.random.random(self._shape).astype(dtype)
+            sc_res = psi(input)
+            for place in self.places:
+                # it is more convenient to use `guard` than `enable/disable_**` here
+                with fluid.dygraph.guard(place):
+                    input_t = paddle.to_tensor(input)
+                    res = paddle.digamma(input_t).numpy()
+                    self.assertEqual(np.allclose(res, sc_res, rtol=1e-05), True)
+
+    def test_name_argument(self):
+        with static.program_guard(static.Program()):
+            x = static.data(name="x", shape=self._shape, dtype=self.dtypes[0])
+            out = paddle.digamma(x, name="digamma_res")
+            self.assertTrue("digamma_res" in out.name)
+
+    def test_dtype_error(self):
+        # in static mode
+        with self.assertRaises(TypeError):
+            with static.program_guard(static.Program()):
+                x = static.data(name="x", shape=self._shape, dtype="int32")
+                out = paddle.digamma(x, name="digamma_res")
+
+        # in dynamic mode
+        with self.assertRaises(RuntimeError):
+            with fluid.dygraph.guard():
+                input = np.random.random(self._shape).astype("int32")
+                input_t = paddle.to_tensor(input)
+                res = paddle.digamma(input_t)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/tensor/__init__.py

@@ -161,6 +161,7 @@
 from .math import any  # noqa: F401
 from .math import broadcast_shape  # noqa: F401
 from .math import conj  # noqa: F401
+from .math import digamma  # noqa: F401
 from .math import neg  # noqa: F401
 from .math import lgamma  # noqa: F401
 
@@ -346,5 +347,6 @@
            'rank',
            'shape',
            'real',
-           'imag'
+           'imag',
+           'digamma'
 ]