add momentum_op_npu and test

Author: ronny1996

paddle/fluid/operators/optimizers/momentum_op_npu.cc

@@ -0,0 +1,56 @@
+/* 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 <string>
+#include "paddle/fluid/operators/npu_op_runner.h"
+#include "paddle/fluid/operators/optimizers/sgd_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class NPUMomentumOpKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto& dev_ctx = ctx.template device_context<platform::NPUDeviceContext>();
+
+    auto param = ctx.Input<framework::Tensor>("Param");
+    auto velocity = ctx.Input<framework::Tensor>("Velocity");
+    auto learning_rate = ctx.Input<framework::Tensor>("LearningRate");
+    auto grad = ctx.Input<framework::Tensor>("Grad");
+    auto param_out = ctx.Output<framework::Tensor>("ParamOut");
+    auto velocity_out = ctx.Output<framework::Tensor>("VelocityOut");
+    param_out->mutable_data<T>(ctx.GetPlace());
+    velocity_out->mutable_data<T>(ctx.GetPlace());
+
+    T mu = static_cast<T>(ctx.Attr<float>("mu"));
+    bool use_nesterov = ctx.Attr<bool>("use_nesterov");
+
+    Tensor mu_tensor;
+    mu_tensor.mutable_data<T>(framework::make_ddim({1}), ctx.GetPlace());
+    FillNpuTensorWithConstant<T>(&mu_tensor, mu);
+    framework::TensorCopy(*param, ctx.GetPlace(), dev_ctx, param_out);
+    framework::TensorCopy(*velocity, ctx.GetPlace(), dev_ctx, velocity_out);
+    const auto& runner = NpuOpRunner(
+        "ApplyMomentum",
+        {*param_out, *velocity_out, *learning_rate, *grad, mu_tensor},
+        {*param_out}, {{"use_nesterov", use_nesterov}});
+    auto stream = dev_ctx.stream();
+    runner.Run(stream);
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_NPU_KERNEL(momentum, ops::NPUMomentumOpKernel<float>);

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

@@ -0,0 +1,88 @@
+#  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.
+
+from __future__ import print_function
+
+import numpy as np
+import unittest
+import sys
+sys.path.append("..")
+from op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+from test_momentum_op import calculate_momentum_by_numpy
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestMomentumOp1(OpTest):
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "momentum"
+        self.init_dtype()
+        self.init_case()
+
+        param = np.random.random(self.shape).astype(self.dtype)
+        grad = np.random.random(self.shape).astype(self.dtype)
+        velocity = np.zeros(self.shape).astype(self.dtype)
+        learning_rate = np.array([0.001]).astype(np.float32)
+        mu = 0.0001
+
+        self.inputs = {
+            'Param': param,
+            'Grad': grad,
+            'Velocity': velocity,
+            'LearningRate': learning_rate
+        }
+
+        self.attrs = {'mu': mu, 'use_nesterov': self.use_nesterov}
+
+        param_out, velocity_out = calculate_momentum_by_numpy(
+            param=param,
+            grad=grad,
+            mu=mu,
+            velocity=velocity,
+            use_nesterov=self.use_nesterov,
+            learning_rate=learning_rate)
+
+        self.outputs = {'ParamOut': param_out, 'VelocityOut': velocity_out}
+
+    def init_case(self):
+        self.shape = (123, 321)
+        self.use_nesterov = False
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def test_check_output(self):
+        self.check_output_with_place(core.NPUPlace(0))
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestMomentumOp2(TestMomentumOp1):
+    def init_case(self):
+        self.shape = (123, 321)
+        self.use_nesterov = True
+
+
+if __name__ == "__main__":
+    unittest.main()