[CustomOp]Add new method for custom double grad (#41538) (#41781)

* add new method for custom double grad

* add tanh double grad unittest

* change year

* revert tensor init method

paddle/fluid/framework/custom_operator.cc

@@ -67,9 +67,17 @@ inline static bool IsDuplicableVar(const std::string& var_name) {
   return var_name.rfind(suffix) != std::string::npos;
 }
 
-inline static std::string NoGrad(const std::string& var_name) {
+inline static std::string NoGrad(const std::string& var_name,
+                                 bool is_double_grad = false) {
   std::string suffix = kGradVarSuffix;
-  return var_name.substr(0, var_name.size() - kGradVarSuffixSize);
+  std::string new_out_suffix = kDoubleGradNewOutSuffix;
+  std::string tmp_var_name(var_name);
+  if (is_double_grad &&
+      (tmp_var_name.rfind(new_out_suffix) != std::string::npos)) {
+    tmp_var_name = tmp_var_name.substr(
+        0, tmp_var_name.size() - /*kDoubleGradNewOutSuffix length*/ 4);
+  }
+  return tmp_var_name.substr(0, tmp_var_name.size() - kGradVarSuffixSize);
 }
 
 inline static bool IsGradVar(const std::string& var_name, bool is_double_grad) {
@@ -533,11 +541,12 @@ class CustomGradOpMaker<OpDesc> : public SingleGradOpMaker<OpDesc> {
     for (auto& out_name : outputs_) {
       VLOG(3) << "Custom Operator: GradOpDescMaker - output: " << out_name;
       if (detail::IsDuplicableVar(out_name)) {
-        grad_op->SetOutput(out_name,
-                           this->InputGrad(detail::NoGrad(out_name),
-                                           /*drop_empty_grad=*/false));
+        grad_op->SetOutput(
+            out_name, this->InputGrad(detail::NoGrad(out_name, is_double_grad_),
+                                      /*drop_empty_grad=*/false));
       } else {
-        grad_op->SetOutput(out_name, this->InputGrad(detail::NoGrad(out_name)));
+        grad_op->SetOutput(out_name, this->InputGrad(detail::NoGrad(
+                                         out_name, is_double_grad_)));
       }
     }
     grad_op->SetAttrMap(this->Attrs());
@@ -600,7 +609,8 @@ class CustomGradOpMaker<imperative::OpBase>
     }
     for (auto& out_name : outputs_) {
       VLOG(3) << "Custom Operator: GradOpBaseMaker - output: " << out_name;
-      grad_op->SetOutput(out_name, this->InputGrad(detail::NoGrad(out_name)));
+      grad_op->SetOutput(
+          out_name, this->InputGrad(detail::NoGrad(out_name, is_double_grad_)));
     }
     grad_op->SetAttrMap(this->Attrs());
   }
@@ -885,8 +895,8 @@ void RegisterOperatorWithMetaInfo(const std::vector<OpMetaInfo>& op_meta_infos,
 
     // Grad InferShape
     if (grad_infer_shape_fn == nullptr) {
-      grad_info.infer_shape_ = [grad_op_inputs,
-                                grad_op_outputs](InferShapeContext* ctx) {
+      grad_info.infer_shape_ = [grad_op_inputs, grad_op_outputs,
+                                is_double_grad](InferShapeContext* ctx) {
         // 1. if forward input exists, gradient's shape is same with forward
         // input
         // default
@@ -897,7 +907,7 @@ void RegisterOperatorWithMetaInfo(const std::vector<OpMetaInfo>& op_meta_infos,
         //    [Suitable for the situation that forward input is not used as
         //    backward input]
         for (auto& out_name : grad_op_outputs) {
-          auto fwd_name = detail::NoGrad(out_name);
+          auto fwd_name = detail::NoGrad(out_name, is_double_grad);
           if (detail::IsDuplicableVar(fwd_name)) {
             // Duplicable forward var must as backward input
             ctx->ShareDim(fwd_name, out_name);

paddle/phi/api/ext/op_meta_info.h

@@ -58,6 +58,7 @@ using Tensor = paddle::experimental::Tensor;
 
 constexpr char kGradTensorSuffix[] = "@GRAD";
 constexpr char kTensorVectorSuffix[] = "@VECTOR";
+constexpr char kDoubleGradNewOutSuffix[] = "@NEW";
 
 // Used for Construct Grad Tensor name
 inline std::string Grad(const std::string& t_name) {
@@ -77,6 +78,15 @@ inline std::string Vec(const std::string& t_name) {
   return result;
 }
 
+// Used for Construct double grad output name
+inline std::string New(const std::string& t_name) {
+  std::string result;
+  result.reserve(t_name.size() + 4U);
+  result += t_name;
+  result += kDoubleGradNewOutSuffix;
+  return result;
+}
+
 PADDLE_API void AssignTensorImpl(const Tensor& src, Tensor* dst);
 
 ////////////////////// Kernel Context ////////////////////////

python/paddle/fluid/tests/custom_op/CMakeLists.txt

@@ -23,6 +23,7 @@ py_test(test_custom_concat SRCS test_custom_concat.py)
 py_test(test_custom_conj SRCS test_custom_conj.py)
 py_test(test_custom_linear SRCS test_custom_linear.py)
 py_test(test_custom_simple_slice SRCS test_custom_simple_slice.py)
+py_test(test_custom_tanh_double_grad SRCS test_custom_tanh_double_grad.py)
 
 # other tests
 py_test(test_sysconfig SRCS test_sysconfig.py)

python/paddle/fluid/tests/custom_op/custom_tanh_op.cc

+// Copyright (c) 2022 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 <cmath>
+#include <iostream>
+#include <vector>
+
+#include "paddle/extension.h"
+
+#define CHECK_CPU_INPUT(x) \
+  PD_CHECK(x.place() == paddle::PlaceType::kCPU, #x " must be a CPU Tensor.")
+
+template <typename data_t>
+void tanh_cpu_forward_kernel(const data_t* x_data,
+                             data_t* out_data,
+                             int64_t x_numel) {
+  PD_CHECK(x_data != nullptr, "x_data is nullptr.");
+  PD_CHECK(out_data != nullptr, "out_data is nullptr.");
+  for (int64_t i = 0; i < x_numel; ++i) {
+    out_data[i] = std::tanh(x_data[i]);
+  }
+}
+
+template <typename data_t>
+void tanh_cpu_backward_kernel(const data_t* grad_out_data,
+                              const data_t* out_data,
+                              data_t* grad_x_data,
+                              int64_t out_numel) {
+  PD_CHECK(grad_out_data != nullptr, "grad_out_data is nullptr.");
+  PD_CHECK(out_data != nullptr, "out_data is nullptr.");
+  PD_CHECK(grad_x_data != nullptr, "grad_x_data is nullptr.");
+  for (int64_t i = 0; i < out_numel; ++i) {
+    grad_x_data[i] =
+        grad_out_data[i] * (static_cast<data_t>(1) - out_data[i] * out_data[i]);
+  }
+}
+
+template <typename data_t>
+void tanh_cpu_double_backward_kernel(const data_t* out_data,
+                                     const data_t* ddx_data,
+                                     const data_t* dout_data,
+                                     data_t* dout_new_data,
+                                     data_t* ddout_data,
+                                     int64_t ddout_numel) {
+  PD_CHECK(out_data != nullptr, "out_data is nullptr.");
+  PD_CHECK(ddx_data != nullptr, "ddx_data is nullptr.");
+  PD_CHECK(dout_data != nullptr, "dout_data is nullptr.");
+  PD_CHECK(dout_new_data != nullptr, "dout_new_data is nullptr.");
+  PD_CHECK(ddout_data != nullptr, "ddout_data is nullptr.");
+  for (int64_t i = 0; i < ddout_numel; ++i) {
+    dout_new_data[i] = static_cast<data_t>(-1) * dout_data[i] *
+                       static_cast<data_t>(2) * out_data[i] * ddx_data[i];
+    ddout_data[i] =
+        ddx_data[i] * (static_cast<data_t>(1) - out_data[i] * out_data[i]);
+  }
+}
+
+std::vector<paddle::Tensor> TanhForward(const paddle::Tensor& x) {
+  CHECK_CPU_INPUT(x);
+  auto out = paddle::Tensor(paddle::PlaceType::kCPU, x.shape());
+
+  PD_DISPATCH_FLOATING_TYPES(
+      x.dtype(), "tanh_cpu_forward", ([&] {
+        tanh_cpu_forward_kernel<data_t>(
+            x.data<data_t>(), out.mutable_data<data_t>(x.place()), x.size());
+      }));
+
+  return {out};
+}
+
+std::vector<paddle::Tensor> TanhBackward(const paddle::Tensor& out,
+                                         const paddle::Tensor& grad_out) {
+  CHECK_CPU_INPUT(out);
+  auto grad_x = paddle::Tensor(paddle::PlaceType::kCPU, out.shape());
+
+  PD_DISPATCH_FLOATING_TYPES(out.dtype(), "tanh_cpu_backward", ([&] {
+                               tanh_cpu_backward_kernel<data_t>(
+                                   grad_out.data<data_t>(),
+                                   out.data<data_t>(),
+                                   grad_x.mutable_data<data_t>(out.place()),
+                                   out.size());
+                             }));
+
+  return {grad_x};
+}
+
+std::vector<paddle::Tensor> TanhDoubleBackward(const paddle::Tensor& out,
+                                               const paddle::Tensor& ddx,
+                                               const paddle::Tensor& dout) {
+  CHECK_CPU_INPUT(out);
+  CHECK_CPU_INPUT(ddx);
+  CHECK_CPU_INPUT(dout);
+  auto dout_new = paddle::Tensor(paddle::PlaceType::kCPU, out.shape());
+  auto ddout = paddle::Tensor(paddle::PlaceType::kCPU, out.shape());
+
+  PD_DISPATCH_FLOATING_TYPES(out.dtype(), "tanh_cpu_double_backward", ([&] {
+                               tanh_cpu_double_backward_kernel<data_t>(
+                                   out.data<data_t>(),
+                                   ddx.data<data_t>(),
+                                   dout.data<data_t>(),
+                                   dout_new.mutable_data<data_t>(out.place()),
+                                   ddout.mutable_data<data_t>(out.place()),
+                                   ddout.size());
+                             }));
+
+  return {dout_new, ddout};
+}
+
+std::vector<std::vector<int64_t>> TanhBackwardInferShape(
+    const std::vector<int64_t>& out_shape,
+    const std::vector<int64_t>& dout_shape) {
+  return {out_shape};
+}
+
+std::vector<std::vector<int64_t>> TanhDoubleBackwardInferShape(
+    const std::vector<int64_t>& out_shape,
+    const std::vector<int64_t>& ddx_shape,
+    const std::vector<int64_t>& dout_shape) {
+  return {dout_shape, dout_shape};
+}
+
+PD_BUILD_OP(custom_tanh)
+    .Inputs({"X"})
+    .Outputs({"Out"})
+    .SetKernelFn(PD_KERNEL(TanhForward));
+
+PD_BUILD_GRAD_OP(custom_tanh)
+    .Inputs({"Out", paddle::Grad("Out")})
+    .Outputs({paddle::Grad("X")})
+    .SetKernelFn(PD_KERNEL(TanhBackward))
+    .SetInferShapeFn(PD_INFER_SHAPE(TanhBackwardInferShape));
+
+PD_BUILD_DOUBLE_GRAD_OP(custom_tanh)
+    .Inputs({"Out", paddle::Grad(paddle::Grad("X")), paddle::Grad("Out")})
+    .Outputs({paddle::New(paddle::Grad("Out")),
+              paddle::Grad(paddle::Grad("Out"))})
+    .SetKernelFn(PD_KERNEL(TanhDoubleBackward))
+    .SetInferShapeFn(PD_INFER_SHAPE(TanhDoubleBackwardInferShape));

python/paddle/fluid/tests/custom_op/test_custom_tanh_double_grad.py

+# Copyright (c) 2022 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 os
+import unittest
+import numpy as np
+
+import paddle
+import paddle.static as static
+from paddle.utils.cpp_extension import load, get_build_directory
+from paddle.utils.cpp_extension.extension_utils import run_cmd
+from utils import paddle_includes, extra_cc_args, extra_nvcc_args
+from paddle.fluid.framework import _test_eager_guard
+
+# Because Windows don't use docker, the shared lib already exists in the
+# cache dir, it will not be compiled again unless the shared lib is removed.
+file = '{}\\custom_tanh\\custom_tanh.pyd'.format(get_build_directory())
+if os.name == 'nt' and os.path.isfile(file):
+    cmd = 'del {}'.format(file)
+    run_cmd(cmd, True)
+
+custom_ops = load(
+    name='custom_tanh_jit',
+    sources=['custom_tanh_op.cc'],
+    extra_include_paths=paddle_includes,  # add for Coverage CI
+    extra_cxx_cflags=extra_cc_args,  # test for cc flags
+    extra_cuda_cflags=extra_nvcc_args,  # test for nvcc flags
+    verbose=True)
+
+
+def custom_tanh_double_grad_dynamic(func, device, dtype, np_x):
+    paddle.set_device(device)
+
+    t = paddle.to_tensor(np_x, dtype=dtype, stop_gradient=False)
+
+    out = func(t)
+    out.stop_gradient = False
+
+    dx = paddle.grad(
+        outputs=[out], inputs=[t], create_graph=True, retain_graph=True)
+
+    dx[0].backward()
+
+    assert out.grad is not None
+    assert dx[0].grad is not None
+    return dx[0].numpy(), dx[0].grad.numpy(), out.grad.numpy()
+
+
+class TestCustomTanhDoubleGradJit(unittest.TestCase):
+    def setUp(self):
+        paddle.set_device('cpu')
+        self.dtypes = ['float32', 'float64']
+        self.devices = ['cpu']
+
+    def test_func_double_grad_dynamic(self):
+        for device in self.devices:
+            for dtype in self.dtypes:
+                x = np.random.uniform(-1, 1, [4, 8]).astype(dtype)
+                out, dx_grad, dout = custom_tanh_double_grad_dynamic(
+                    custom_ops.custom_tanh, device, dtype, x)
+                pd_out, pd_dx_grad, pd_dout = custom_tanh_double_grad_dynamic(
+                    paddle.tanh, device, dtype, x)
+                self.assertTrue(
+                    np.allclose(out, pd_out),
+                    "custom op out: {},\n paddle api out: {}".format(out,
+                                                                     pd_out))
+                self.assertTrue(
+                    np.allclose(dx_grad, pd_dx_grad),
+                    "custom op dx grad: {},\n paddle api dx grad: {}".format(
+                        dx_grad, pd_dx_grad))
+                self.assertTrue(
+                    np.allclose(dout, pd_dout),
+                    "custom op out grad: {},\n paddle api out grad: {}".format(
+                        dout, pd_dout))
+
+
+if __name__ == "__main__":
+    unittest.main()