• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2021 Huawei Technologies Co., Ltd
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14
15import numpy as np
16
17import mindspore as ms
18from mindspore import context, Tensor, Parameter
19from mindspore.common.api import _cell_graph_executor
20from mindspore.nn import Cell, TrainOneStepCell, Momentum, BatchNorm2d, BatchNorm1d
21from mindspore.ops import operations as P
22
23
24class Net(Cell):
25    def __init__(self, conv2d_weight, out_channel, kernel_size, pad_mode, stride,
26                 strategy1=None, strategy2=None):
27        super().__init__()
28        self.conv2d = P.Conv2D(out_channel=out_channel, kernel_size=kernel_size,
29                               pad_mode=pad_mode, stride=stride).shard(strategy1)
30        self.conv2d_weight = Parameter(conv2d_weight, "w1")
31        self.bn = BatchNorm2d(8)
32        self.bn.bn_train.shard(strategy2)
33
34    def construct(self, x, b):
35        out = self.conv2d(x, self.conv2d_weight)
36        out = self.bn(out)
37        return out
38
39
40_x = Tensor(np.ones([32, 16, 8, 8]), dtype=ms.float32)
41_w1 = Tensor(np.ones([8, 16, 2, 2]), dtype=ms.float32)
42_b = Tensor(np.ones([32, 16, 8, 8]), dtype=ms.float32)
43
44
45def compile_net(net):
46    optimizer = Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
47    train_net = TrainOneStepCell(net, optimizer)
48    train_net.set_auto_parallel()
49    train_net.set_train()
50    _cell_graph_executor.compile(train_net, _x, _b)
51    context.reset_auto_parallel_context()
52
53
54def test_batchnorm_data_parallel():
55    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
56    strategy1 = ((8, 1, 1, 1), (1, 1, 1, 1))
57    strategy2 = ((8, 1, 1, 1), (1,), (1,), (1,), (1,))
58    net = Net(_w1, out_channel=8, kernel_size=2, pad_mode="same", stride=1, strategy1=strategy1, strategy2=strategy2)
59    compile_net(net)
60
61
62def test_batchnorm_model_parallel1():
63    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
64    strategy1 = ((2, 2, 1, 1), (2, 2, 1, 1))
65    strategy2 = ((2, 1, 2, 2), (1,), (1,), (1,), (1,))
66    net = Net(_w1, out_channel=8, kernel_size=2, pad_mode="same", stride=1, strategy1=strategy1, strategy2=strategy2)
67    compile_net(net)
68
69
70def test_batchnorm_model_parallel2():
71    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=32, global_rank=0)
72    strategy1 = ((2, 2, 2, 2), (2, 2, 1, 1))
73    strategy2 = ((1, 8, 1, 1), (8,), (8,), (8,), (8,))
74    net = Net(_w1, out_channel=8, kernel_size=2, pad_mode="same", stride=2, strategy1=strategy1, strategy2=strategy2)
75    compile_net(net)
76
77
78class Net2(Cell):
79    def __init__(self, strategy1=None, strategy2=None):
80        super().__init__()
81        self.bn = BatchNorm1d(8)
82        self.bn.bn_train.shard(strategy1)
83        self.relu = P.ReLU().shard(strategy2)
84
85    def construct(self, x, b):
86        out = self.bn(x)
87        out = self.relu(out)
88        return out
89
90
91_x1 = Tensor(np.ones([32, 8]), dtype=ms.float32)
92_b1 = Tensor(np.ones([32, 8]), dtype=ms.float32)
93
94
95def compile_net2(net):
96    optimizer = Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
97    train_net = TrainOneStepCell(net, optimizer)
98    train_net.set_auto_parallel()
99    train_net.set_train()
100    _cell_graph_executor.compile(train_net, _x1, _b1)
101    context.reset_auto_parallel_context()
102
103
104def test_batchnorm1d_data_parallel():
105    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
106    strategy1 = ((8, 1), (1,), (1,), (1,), (1,))
107    strategy2 = ((8, 1),)
108    net = Net2(strategy1=strategy1, strategy2=strategy2)
109    compile_net2(net)
110
111
112def test_batchnorm1d_model_parallel1():
113    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
114    strategy1 = ((1, 8), (8,), (8,), (8,), (8,))
115    strategy2 = ((1, 8),)
116    net = Net2(strategy1=strategy1, strategy2=strategy2)
117    compile_net2(net)
118
119
120def test_batchnorm1d_model_parallel2():
121    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=32, global_rank=0)
122    strategy1 = ((2, 4), (4,), (4,), (4,), (4,))
123    strategy2 = ((2, 4),)
124    net = Net2(strategy1=strategy1, strategy2=strategy2)
125    compile_net2(net)
126