yangzhenzhang
4 years ago
parent
d79bcc923e
commit
0a79ab82ae
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,76 @@
|
||||
/**
|
||||
* Copyright 2020 Huawei Technologies Co., Ltd
|
||||
*
|
||||
* 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.
|
||||
*/
|
||||
|
||||
#ifndef MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_TENSORDOT_INFO_H_
|
||||
#define MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_TENSORDOT_INFO_H_
|
||||
|
||||
#include <memory>
|
||||
#include <string>
|
||||
#include <unordered_map>
|
||||
#include <vector>
|
||||
|
||||
#include "utils/ms_utils.h"
|
||||
#include "ir/value.h"
|
||||
#include "frontend/parallel/auto_parallel/operator_costmodel.h"
|
||||
#include "frontend/parallel/ops_info/operator_info.h"
|
||||
#include "frontend/parallel/strategy.h"
|
||||
#include "frontend/parallel/tensor_layout/tensor_redistribution.h"
|
||||
|
||||
namespace mindspore {
|
||||
namespace parallel {
|
||||
enum AxesType {
|
||||
INT_TYPE = 0,
|
||||
TUPLE_TYPE,
|
||||
TUPLE_TUPLE_TYPE,
|
||||
};
|
||||
|
||||
class TensorDotInfo : public OperatorInfo {
|
||||
public:
|
||||
TensorDotInfo(const std::string &name, const Shapes &inputs_shape, const Shapes &outputs_shape,
|
||||
const PrimitiveAttrs &attrs)
|
||||
: OperatorInfo(name, inputs_shape, outputs_shape, attrs, std::make_shared<MatMulCost>(true)) {}
|
||||
~TensorDotInfo() override = default;
|
||||
|
||||
Status Init(const StrategyPtr &strategy) override;
|
||||
Status InitForCostModel(const StrategyPtr &strategy) override;
|
||||
|
||||
Status GenerateStrategies(int64_t stage_id) override;
|
||||
Status SetCostUnderStrategy(const StrategyPtr &strategy) override;
|
||||
Status PrepareStrategy(int32_t stage_id, size_t dev_num, Dimensions combined_partitions, size_t input0_shape_size,
|
||||
size_t input1_shape_size, StrategyPtr *sp);
|
||||
|
||||
protected:
|
||||
Status CheckStrategy(const StrategyPtr &strategy) override;
|
||||
Status InferMirrorOps() override;
|
||||
Status InferForwardCommunication() override;
|
||||
Status InferTensorInfo() override;
|
||||
Status InferDevMatrixShape() override;
|
||||
Status InferTensorMap() override;
|
||||
Status GetAttrs() override;
|
||||
std::shared_ptr<Strategys> GenerateBatchStrategies() override;
|
||||
void InferTensorMapAxesInt(const TensorMap &tensor_map_index);
|
||||
void InferTensorMapAxesTuple(size_t size, const TensorMap &input_a_tensor_map, const TensorMap &tensor_map_index);
|
||||
void ShowAxes();
|
||||
Shape origin_dev_matrix_shape_;
|
||||
|
||||
AxesType axes_type_ = INT_TYPE;
|
||||
int32_t axes_int_ = 1;
|
||||
std::vector<int32_t> axes_tuple_;
|
||||
std::vector<std::vector<int32_t>> axes_tuple_tuple_;
|
||||
};
|
||||
} // namespace parallel
|
||||
} // namespace mindspore
|
||||
#endif // MINDSPORE_CCSRC_FRONTEND_PARALLEL_OPS_INFO_TENSORDOT_INFO_H_
|
||||
@ -0,0 +1,86 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# 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 numpy as np
|
||||
|
||||
import mindspore as ms
|
||||
from mindspore import context, Tensor, Parameter
|
||||
from mindspore.common.api import _executor
|
||||
from mindspore.nn import Cell, TrainOneStepCell, Momentum
|
||||
from mindspore.ops import operations as P
|
||||
|
||||
|
||||
class Net(Cell):
|
||||
def __init__(self, mul_weight, strategy1=None, strategy2=None):
|
||||
super().__init__()
|
||||
self.mul = P.Mul().shard(strategy1)
|
||||
self.repeat = P.RepeatElements(rep=2, axis=1).shard(strategy2)
|
||||
self.mul_weight = Parameter(mul_weight, "w1")
|
||||
|
||||
def construct(self, x, b):
|
||||
out = self.mul(x, self.mul_weight)
|
||||
out = self.repeat(out)
|
||||
return out
|
||||
|
||||
|
||||
_x = Tensor(np.ones([128, 64, 32]), dtype=ms.float32)
|
||||
_w1 = Tensor(np.ones([128, 64, 32]), dtype=ms.float32)
|
||||
_b = Tensor(np.ones([128, 64, 32]), dtype=ms.float32)
|
||||
|
||||
|
||||
def compile_net(net):
|
||||
optimizer = Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
|
||||
train_net = TrainOneStepCell(net, optimizer)
|
||||
train_net.set_auto_parallel()
|
||||
train_net.set_train()
|
||||
_executor.compile(train_net, _x, _b)
|
||||
context.reset_auto_parallel_context()
|
||||
|
||||
|
||||
def test_repeat_elements_data_parallel():
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
|
||||
strategy1 = ((16, 1, 1), (16, 1, 1))
|
||||
strategy2 = ((16, 1, 1),)
|
||||
net = Net(_w1, strategy1, strategy2)
|
||||
compile_net(net)
|
||||
|
||||
|
||||
def test_repeat_elements_model_parallel():
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
|
||||
strategy1 = ((1, 1, 16), (1, 1, 16))
|
||||
strategy2 = ((1, 1, 16),)
|
||||
net = Net(_w1, strategy1, strategy2)
|
||||
compile_net(net)
|
||||
|
||||
|
||||
def test_repeat_elements_hybrid_parallel():
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
|
||||
strategy1 = ((2, 2, 4), (2, 2, 4))
|
||||
strategy2 = ((2, 2, 4),)
|
||||
net = Net(_w1, strategy1, strategy2)
|
||||
compile_net(net)
|
||||
|
||||
|
||||
def test_repeat_elements_auto_parallel():
|
||||
context.set_auto_parallel_context(parallel_mode="auto_parallel", device_num=16, global_rank=0)
|
||||
net = Net(_w1)
|
||||
compile_net(net)
|
||||
|
||||
|
||||
def test_repeat_elements_repeat_calc():
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
|
||||
strategy1 = ((2, 2, 4), (2, 2, 4))
|
||||
strategy2 = ((1, 2, 2),)
|
||||
net = Net(_w1, strategy1, strategy2)
|
||||
compile_net(net)
|
||||
Loading…
Reference in new issue