From d710c3a04522031ba1273f6d64a31a789f09ca10 Mon Sep 17 00:00:00 2001 From: ronnywang <524019753@qq.com> Date: Wed, 25 Aug 2021 10:20:01 +0800 Subject: [PATCH] [NPU] add npu_one_hot_v2 (#34937) --- paddle/fluid/operators/one_hot_v2_op_npu.cc | 81 ++++++ .../unittests/npu/test_one_hot_v2_op_npu.py | 240 ++++++++++++++++++ 2 files changed, 321 insertions(+) create mode 100644 paddle/fluid/operators/one_hot_v2_op_npu.cc create mode 100644 python/paddle/fluid/tests/unittests/npu/test_one_hot_v2_op_npu.py diff --git a/paddle/fluid/operators/one_hot_v2_op_npu.cc b/paddle/fluid/operators/one_hot_v2_op_npu.cc new file mode 100644 index 00000000000..9a25b8e5224 --- /dev/null +++ b/paddle/fluid/operators/one_hot_v2_op_npu.cc @@ -0,0 +1,81 @@ +/* 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/one_hot_v2_op.h" + +#include "paddle/fluid/operators/npu_op_runner.h" + +namespace paddle { +namespace operators { +using Tensor = framework::Tensor; + +template +class OneHotV2NPUKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& dev_ctx = + ctx.template device_context(); + auto* in = ctx.Input("X"); + auto* out = ctx.Output("Out"); + int depth = ctx.Attr("depth"); + + if (ctx.HasInput("depth_tensor")) { + auto* depth_tensor = ctx.Input("depth_tensor"); + std::vector depth_data; + framework::TensorToVector(*depth_tensor, dev_ctx, &depth_data); + depth = depth_data[0]; + auto out_dims = out->dims(); + out_dims[out_dims.size() - 1] = depth; + out->Resize(out_dims); + } + out->mutable_data(ctx.GetPlace()); + + float on_value = 1.0f, off_value = 0.0f; + if (in->type() == framework::proto::VarType::INT32) { + NpuOpRunner runner; + runner.SetType("OneHot") + .AddInput(*in) + .AddInput(std::vector({static_cast(depth)})) + .AddInput(std::vector({on_value})) + .AddInput(std::vector({off_value})) + .AddAttr("axis", -1) + .AddOutput(*out); + runner.Run(dev_ctx.stream()); + } else { + Tensor transformed_in; + transformed_in.mutable_data(in->dims(), dev_ctx.GetPlace()); + const auto& cast_runner = NpuOpRunner("Cast", {*in}, {transformed_in}, + {{"dst_type", ACL_INT32}}); + cast_runner.Run(dev_ctx.stream()); + NpuOpRunner runner; + runner.SetType("OneHot") + .AddInput(transformed_in) + .AddInput(std::vector({static_cast(depth)})) + .AddInput(std::vector({on_value})) + .AddInput(std::vector({off_value})) + .AddAttr("axis", -1) + .AddOutput(*out); + runner.Run(dev_ctx.stream()); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +namespace plat = paddle::platform; + +REGISTER_OP_NPU_KERNEL(one_hot_v2, ops::OneHotV2NPUKernel, + ops::OneHotV2NPUKernel); diff --git a/python/paddle/fluid/tests/unittests/npu/test_one_hot_v2_op_npu.py b/python/paddle/fluid/tests/unittests/npu/test_one_hot_v2_op_npu.py new file mode 100644 index 00000000000..e511286cc2d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/npu/test_one_hot_v2_op_npu.py @@ -0,0 +1,240 @@ +# 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 sys +import unittest +import numpy as np +sys.path.append("..") + +from op_test import OpTest +import paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.fluid.framework import Program, program_guard + +paddle.enable_static() + + +class TestOneHotOp(OpTest): + def set_npu(self): + self.__class__.use_npu = True + + def setUp(self): + self.set_npu() + self.op_type = 'one_hot_v2' + depth = 10 + depth_np = np.array(10).astype('int32') + dimension = 12 + x_lod = [[4, 1, 3, 3]] + x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))] + x = np.array(x).astype('int32').reshape([sum(x_lod[0])]) + + out = np.zeros(shape=(np.product(x.shape), depth)).astype('float32') + + for i in range(np.product(x.shape)): + out[i, x[i]] = 1.0 + + self.inputs = {'X': (x, x_lod), 'depth_tensor': depth_np} + self.attrs = {'dtype': int(core.VarDesc.VarType.FP32)} + self.outputs = {'Out': (out, x_lod)} + + def test_check_output(self): + self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False) + + +class TestOneHotOp_non_lod(OpTest): + def setUp(self): + self.op_type = 'one_hot_v2' + depth = 10 + depth_np = np.array(10).astype('int32') + dimension = 12 + x_lod = [[4, 1, 3, 3]] + x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))] + x = np.array(x).astype('int32').reshape([sum(x_lod[0])]) + + out = np.zeros(shape=(np.product(x.shape), depth)).astype('float32') + + for i in range(np.product(x.shape)): + out[i, x[i]] = 1.0 + + self.inputs = {'X': x, 'depth_tensor': depth_np} + self.attrs = {'dtype': int(core.VarDesc.VarType.FP32)} + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + +class TestOneHotOp_attr(OpTest): + def set_npu(self): + self.__class__.use_npu = True + + def setUp(self): + self.set_npu() + self.op_type = 'one_hot_v2' + depth = 10 + dimension = 12 + x_lod = [[4, 1, 3, 3]] + x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))] + x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1]) + + out = np.zeros(shape=(np.product(x.shape[:-1]), 1, + depth)).astype('float32') + + for i in range(np.product(x.shape)): + out[i, 0, x[i]] = 1.0 + + self.inputs = {'X': (x, x_lod)} + self.attrs = {'dtype': int(core.VarDesc.VarType.FP32), 'depth': depth} + self.outputs = {'Out': (out, x_lod)} + + def test_check_output(self): + self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False) + + +class TestOneHotOp_default_dtype(OpTest): + def set_npu(self): + self.__class__.use_npu = True + + def setUp(self): + self.set_npu() + self.op_type = 'one_hot_v2' + depth = 10 + depth_np = np.array(10).astype('int32') + dimension = 12 + x_lod = [[4, 1, 3, 3]] + x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))] + x = np.array(x).astype('int32').reshape([sum(x_lod[0])]) + + out = np.zeros(shape=(np.product(x.shape), depth)).astype('float32') + + for i in range(np.product(x.shape)): + out[i, x[i]] = 1.0 + + self.inputs = {'X': (x, x_lod), 'depth_tensor': depth_np} + self.attrs = {} + self.outputs = {'Out': (out, x_lod)} + + def test_check_output(self): + self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False) + + +class TestOneHotOp_default_dtype_attr(OpTest): + def set_npu(self): + self.__class__.use_npu = True + + def setUp(self): + self.set_npu() + self.op_type = 'one_hot_v2' + depth = 10 + dimension = 12 + x_lod = [[4, 1, 3, 3]] + x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))] + x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1]) + + out = np.zeros(shape=(np.product(x.shape[:-1]), 1, + depth)).astype('float32') + + for i in range(np.product(x.shape)): + out[i, 0, x[i]] = 1.0 + + self.inputs = {'X': (x, x_lod)} + self.attrs = {'depth': depth} + self.outputs = {'Out': (out, x_lod)} + + def test_check_output(self): + self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False) + + +class TestOneHotOp_out_of_range(OpTest): + def set_npu(self): + self.__class__.use_npu = True + + def setUp(self): + self.set_npu() + self.op_type = 'one_hot_v2' + depth = 10 + x_lod = [[4, 1, 3, 3]] + x = [np.random.choice([-1, depth]) for i in range(sum(x_lod[0]))] + x = np.array(x).astype('int32').reshape([sum(x_lod[0])]) + + out = np.zeros(shape=(np.product(x.shape), depth)).astype('float32') + + self.inputs = {'X': (x, x_lod)} + self.attrs = {'depth': depth, 'allow_out_of_range': True} + self.outputs = {'Out': (out, x_lod)} + + def test_check_output(self): + self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False) + + +class TestOneHotOp_dtype_int64(OpTest): + def set_npu(self): + self.__class__.use_npu = True + + def setUp(self): + self.set_npu() + self.op_type = 'one_hot_v2' + depth = 10 + x_lod = [[4, 1, 3, 3]] + x = [np.random.choice([-1, depth]) for i in range(sum(x_lod[0]))] + x = np.array(x).astype('int64').reshape([sum(x_lod[0])]) + + out = np.zeros(shape=(np.product(x.shape), depth)).astype('float32') + + self.inputs = {'X': (x, x_lod)} + self.attrs = {'depth': depth, 'allow_out_of_range': True} + self.outputs = {'Out': (out, x_lod)} + + def test_check_output(self): + self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False) + + +class TestOneHotOpApi(unittest.TestCase): + def test_api(self): + depth = 10 + self._run(depth) + + def test_api_with_depthTensor(self): + depth = fluid.layers.assign(input=np.array([10], dtype=np.int32)) + self._run(depth) + + def test_api_with_dygraph(self): + depth = 10 + label = np.array([np.random.randint(0, depth - 1) + for i in range(6)]).reshape([6, 1]) + with fluid.dygraph.guard(paddle.NPUPlace(0)): + one_hot_label = fluid.one_hot( + input=fluid.dygraph.to_variable(label), depth=depth) + + def _run(self, depth): + label = fluid.layers.data(name="label", shape=[1], dtype="int64") + one_hot_label = fluid.one_hot(input=label, depth=depth) + + place = fluid.NPUPlace(0) + label_data = np.array([np.random.randint(0, 10 - 1) + for i in range(6)]).reshape([6, 1]) + + exe = fluid.Executor(place) + exe.run(fluid.default_startup_program()) + ret = exe.run(feed={'label': label_data, }, + fetch_list=[one_hot_label], + return_numpy=False) + + +if __name__ == '__main__': + paddle.enable_static() + unittest.main() -- GitLab