From a268c7ce5c8c502164c65e05912e4405c7c149af Mon Sep 17 00:00:00 2001 From: TTerror Date: Wed, 5 Jan 2022 10:34:46 +0800 Subject: [PATCH] add huber_loss for kunlun (#38589) * add huber_loss for kunlun * update xpu.cmake * update unitests * update unitests * update elementwise_add * update elementwise_add * update elementwise_add --- paddle/fluid/operators/huber_loss_op_xpu.cc | 92 +++++++++++++++ .../fluid/platform/device/xpu/xpu2_op_list.h | 3 + .../unittests/xpu/test_huber_loss_op_xpu.py | 110 ++++++++++++++++++ 3 files changed, 205 insertions(+) create mode 100644 paddle/fluid/operators/huber_loss_op_xpu.cc create mode 100644 python/paddle/fluid/tests/unittests/xpu/test_huber_loss_op_xpu.py diff --git a/paddle/fluid/operators/huber_loss_op_xpu.cc b/paddle/fluid/operators/huber_loss_op_xpu.cc new file mode 100644 index 00000000000..767ce542736 --- /dev/null +++ b/paddle/fluid/operators/huber_loss_op_xpu.cc @@ -0,0 +1,92 @@ +/* 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. */ + +#ifdef PADDLE_WITH_XPU + +#include "paddle/fluid/operators/huber_loss_op.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class HuberLossXPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in0 = ctx.Input("X"); + auto* in1 = ctx.Input("Y"); + auto* residual = ctx.Output("Residual"); + auto* out = ctx.Output("Out"); + auto delta = ctx.Attr("delta"); + + auto residual_data = residual->mutable_data(ctx.GetPlace()); + auto out_data = out->mutable_data(ctx.GetPlace()); + auto in0_data = in0->data(); + auto in1_data = in1->data(); + + auto& dev_ctx = + ctx.template device_context(); + int r = xpu::huber_loss(dev_ctx.x_context(), in0_data, in1_data, + residual_data, out_data, in0->numel(), 1, delta); + PADDLE_ENFORCE_EQ(r, XPU_SUCCESS, platform::errors::External( + "XPU API(huber_loss) return wrong " + "value[%d %s]", + r, XPUAPIErrorMsg[r])); + } +}; + +template +class HuberLossGradXPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* residual = ctx.Input("Residual"); + auto* dout = ctx.Input(framework::GradVarName("Out")); + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dy = ctx.Output(framework::GradVarName("Y")); + auto delta = ctx.Attr("delta"); + + T* dx_data = nullptr; + T* dy_data = nullptr; + if (dx) { + dx_data = dx->mutable_data(ctx.GetPlace()); + } + if (dy) { + dy_data = dy->mutable_data(ctx.GetPlace()); + } + auto dout_data = dout->data(); + auto residual_data = residual->data(); + auto& dev_ctx = + ctx.template device_context(); + int r = + xpu::huber_loss_grad(dev_ctx.x_context(), residual_data, dout_data, + dx_data, dy_data, dout->numel(), 1, delta); + PADDLE_ENFORCE_EQ( + r, XPU_SUCCESS, + platform::errors::External("XPU API(huber_loss_grad) return wrong " + "value[%d %s]", + r, XPUAPIErrorMsg[r])); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +namespace plat = paddle::platform; + +REGISTER_OP_XPU_KERNEL(huber_loss, ops::HuberLossXPUKernel); +REGISTER_OP_XPU_KERNEL(huber_loss_grad, ops::HuberLossGradXPUKernel); + +#endif diff --git a/paddle/fluid/platform/device/xpu/xpu2_op_list.h b/paddle/fluid/platform/device/xpu/xpu2_op_list.h index 142685a64ef..79261a5d7bc 100644 --- a/paddle/fluid/platform/device/xpu/xpu2_op_list.h +++ b/paddle/fluid/platform/device/xpu/xpu2_op_list.h @@ -192,6 +192,9 @@ XPUOpMap& get_kl2_ops() { {"hard_swish_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()), pOpKernelType(vartype::FP16, XPUPlace())})}, + {"huber_loss_grad", + XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})}, + {"huber_loss", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})}, {"iou_similarity", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})}, {"label_smooth", diff --git a/python/paddle/fluid/tests/unittests/xpu/test_huber_loss_op_xpu.py b/python/paddle/fluid/tests/unittests/xpu/test_huber_loss_op_xpu.py new file mode 100644 index 00000000000..0cd98d2daea --- /dev/null +++ b/python/paddle/fluid/tests/unittests/xpu/test_huber_loss_op_xpu.py @@ -0,0 +1,110 @@ +# Copyright (c) 2018 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. + +from __future__ import print_function + +import unittest +import numpy as np +import sys +sys.path.append("..") +from op_test import OpTest +from op_test_xpu import XPUOpTest +import paddle +import paddle.fluid as fluid +from paddle.fluid import compiler, Program, program_guard + +paddle.enable_static() + + +def huber_loss_forward(val, delta): + abs_val = abs(val) + if abs_val <= delta: + return 0.5 * val * val + else: + return delta * (abs_val - 0.5 * delta) + + +class TestHuberLossOp(XPUOpTest): + def setUp(self): + self.set_xpu() + self.op_type = 'huber_loss' + self.place = paddle.XPUPlace(0) + + self.init_dtype() + + self.set_inputs() + self.set_attrs() + self.set_outputs() + + def set_inputs(self): + shape = self.set_shape() + x = np.random.uniform(0, 1., shape).astype(self.dtype) + y = np.random.uniform(0, 1., shape).astype(self.dtype) + self.inputs = { + 'X': OpTest.np_dtype_to_fluid_dtype(x), + 'Y': OpTest.np_dtype_to_fluid_dtype(y) + } + + def set_attrs(self): + self.attrs = {'delta': 0.5} + + def set_outputs(self): + delta = self.attrs['delta'] + shape = self.set_shape() + residual = self.inputs['Y'] - self.inputs['X'] + loss = np.vectorize(huber_loss_forward)(residual, + delta).astype(self.dtype) + self.outputs = {'Residual': residual, 'Out': loss.reshape(shape)} + + def set_shape(self): + return (100, 1) + + def set_xpu(self): + self.__class__.use_xpu = True + + def init_dtype(self): + self.dtype = np.float32 + + def test_check_output(self): + self.check_output_with_place(self.place) + + def test_check_grad_normal(self): + self.check_grad_with_place(self.place, ['X', 'Y'], 'Out') + + def test_check_grad_ingore_x(self): + self.check_grad_with_place( + self.place, ['Y'], 'Out', no_grad_set=set("residual")) + + def test_check_grad_ingore_y(self): + self.check_grad_with_place( + self.place, ['X'], 'Out', no_grad_set=set('residual')) + + +def TestHuberLossOp1(TestHuberLossOp): + def set_shape(self): + return (64) + + +def TestHuberLossOp2(TestHuberLossOp): + def set_shape(self): + return (6, 6) + + +def TestHuberLossOp3(TestHuberLossOp): + def set_shape(self): + return (6, 6, 1) + + +if __name__ == '__main__': + unittest.main() -- GitLab