[NPU] Add huber_loss op (#34826)

* [NPU] Add huber_loss op

* [NPU] Add huber_loss op

* [NPU] Add huber_loss p[

* [NPU] Add huber_loss

paddle/fluid/operators/huber_loss_op_npu.cc

+/* 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 Licnse. */
+
+#include "paddle/fluid/operators/huber_loss_op.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+void HuberLossSub(const platform::Place& place, const aclrtStream& stream,
+                  const Tensor* x, const Tensor* y, Tensor* z) {
+  //  Calculate z = x - y
+  z->mutable_data<T>(x->dims(), place);
+  const auto& runner = NpuOpRunner("Sub", {*x, *y}, {*z}, {});
+  runner.Run(stream);
+}
+
+template <typename T>
+void HuberLossMuls(const platform::Place& place, const aclrtStream& stream,
+                   const Tensor* x, float scalar, Tensor* y) {
+  //  Calculate y = x + scale
+  y->mutable_data<T>(x->dims(), place);
+  const auto& runner = NpuOpRunner("Muls", {*x}, {*y}, {{"value", scalar}});
+  runner.Run(stream);
+}
+
+template <typename T>
+void HuberLossZerosLike(const platform::Place& place, const aclrtStream& stream,
+                        const Tensor* x, Tensor* y) {
+  y->mutable_data<T>(x->dims(), place);
+  const auto& runner = NpuOpRunner("ZerosLike", {*x}, {*y}, {});
+  runner.Run(stream);
+}
+
+template <typename T>
+void HuberLossSmoothL1Loss(const platform::Place& place,
+                           const aclrtStream& stream, const Tensor* x,
+                           const Tensor* y, float delta, Tensor* z) {
+  z->mutable_data<T>(x->dims(), place);
+  const auto& runner =
+      NpuOpRunner("SmoothL1Loss", {*x, *y}, {*z}, {{"sigma", delta}});
+  runner.Run(stream);
+}
+
+template <typename T>
+void HuberLossSmoothL1LossGrad(const platform::Place& place,
+                               const aclrtStream& stream, const Tensor* pred,
+                               const Tensor* lab, const Tensor* dout,
+                               float sigma, Tensor* grad) {
+  grad->mutable_data<T>(pred->dims(), place);
+  const auto& runner = NpuOpRunner("SmoothL1LossGrad", {*pred, *lab, *dout},
+                                   {*grad}, {{"sigma", sigma}});
+  runner.Run(stream);
+}
+
+template <typename T>
+class HuberLossNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in0 = ctx.Input<Tensor>("X");
+    auto* in1 = ctx.Input<Tensor>("Y");
+    auto* residual = ctx.Output<Tensor>("Residual");
+    auto* out = ctx.Output<Tensor>("Out");
+    auto delta = ctx.Attr<float>("delta");
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    auto place = ctx.GetPlace();
+    HuberLossSub<T>(place, stream, in1, in0, residual);
+
+    HuberLossSmoothL1Loss<T>(place, stream, in0, in1, delta, out);
+    HuberLossMuls<T>(place, stream, out, delta, out);
+  }
+};
+
+template <typename T>
+class HuberLossGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* residual = ctx.Input<Tensor>("Residual");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto delta = ctx.Attr<float>("delta");
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    auto place = ctx.GetPlace();
+
+    Tensor t_grad_rd;
+    if (dx || dy) {
+      Tensor t_zero;
+      HuberLossZerosLike<T>(place, stream, residual, &t_zero);
+      HuberLossSmoothL1LossGrad<T>(place, stream, residual, &t_zero, dout,
+                                   delta, &t_grad_rd);
+    }
+    if (dx) {
+      HuberLossMuls<T>(place, stream, &t_grad_rd, -delta, dx);
+    }
+    if (dy) {
+      HuberLossMuls<T>(place, stream, &t_grad_rd, delta, dy);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
+REGISTER_OP_NPU_KERNEL(huber_loss, ops::HuberLossNPUKernel<float>,
+                       ops::HuberLossNPUKernel<plat::float16>);
+REGISTER_OP_NPU_KERNEL(huber_loss_grad, ops::HuberLossGradNPUKernel<float>,
+                       ops::HuberLossGradNPUKernel<plat::float16>);

python/paddle/fluid/tests/unittests/npu/test_huber_loss_op_npu.py

+#   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
+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)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestHuberLossOp(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = 'huber_loss'
+        self.place = paddle.NPUPlace(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_npu(self):
+        self.__class__.use_npu = 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):
+        if self.dtype == np.float16:
+            return
+        self.check_grad_with_place(self.place, ['X', 'Y'], 'Out')
+
+    def test_check_grad_ingore_x(self):
+        if self.dtype == np.float16:
+            return
+        self.check_grad_with_place(
+            self.place, ['Y'],
+            'Out',
+            max_relative_error=0.008,
+            no_grad_set=set("residual"))
+
+    def test_check_grad_ingore_y(self):
+        if self.dtype == np.float16:
+            return
+        self.check_grad_with_place(
+            self.place, ['X'],
+            'Out',
+            max_relative_error=0.008,
+            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)
+
+
+def TestHuberLossOpFP16(TestHuberLossOp):
+    def init_dtype(self):
+        self.dtype = np.float16
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestHuberLossOpError(unittest.TestCase):
+    def test_errors(self):
+        with program_guard(Program(), Program()):
+            # the input and label must be Variable
+            xw = np.random.random((6, 6)).astype("float32")
+            xr = fluid.data(name='xr', shape=[None, 6], dtype="float32")
+            lw = np.random.random((6, 6)).astype("float32")
+            lr = fluid.data(name='lr', shape=[None, 6], dtype="float32")
+            delta = 1.0
+            self.assertRaises(TypeError, fluid.layers.huber_loss, xr, lw, delta)
+            self.assertRaises(TypeError, fluid.layers.huber_loss, xw, lr, delta)
+
+            # the dtype of input and label must be float32 or float64
+            xw2 = fluid.data(name='xw2', shape=[None, 6], dtype="int32")
+            lw2 = fluid.data(name='lw2', shape=[None, 6], dtype="int32")
+            self.assertRaises(TypeError, fluid.layers.huber_loss, xw2, lr,
+                              delta)
+            self.assertRaises(TypeError, fluid.layers.huber_loss, xr, lw2,
+                              delta)
+
+
+if __name__ == '__main__':
+    unittest.main()