• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2019 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
18import mindspore.nn as nn
19from mindspore import Tensor
20from mindspore import context
21from mindspore.common.api import _cell_graph_executor
22from mindspore.ops import composite as C
23from mindspore.ops import operations as P
24from tests.ut.python.ops.test_math_ops import VirtualLoss
25
26
27grad_all = C.GradOperation(get_all=True)
28
29
30class NetWithLoss(nn.Cell):
31    def __init__(self, network):
32        super(NetWithLoss, self).__init__()
33        self.loss = VirtualLoss()
34        self.network = network
35
36    def construct(self, x, w1, w2):
37        predict = self.network(x, w1, w2)
38        return self.loss(predict)
39
40
41class GradWrap(nn.Cell):
42    def __init__(self, network):
43        super(GradWrap, self).__init__()
44        self.network = network
45
46    def construct(self, x, w1, w2):
47        return grad_all(self.network)(x, w1, w2)
48
49
50class NetConv(nn.Cell):
51    def __init__(self,
52                 cin,
53                 cout,
54                 kernel_size,
55                 stride=1,
56                 pad_mode='pad',
57                 padding=0,
58                 dilation=1,
59                 group=1,
60                 has_bias=False,
61                 weight_init='normal',
62                 bias_init='zeros',
63                 strategy=None):
64        super(NetConv, self).__init__()
65        self.conv = nn.Conv2d(cin,
66                              cout,
67                              kernel_size,
68                              stride,
69                              pad_mode,
70                              padding,
71                              dilation,
72                              group,
73                              has_bias,
74                              weight_init,
75                              bias_init)
76        self.conv.conv2d.shard(strategy)
77
78    def construct(self, input_x):
79        return self.conv(input_x)
80
81
82def test_batch():
83    class Net(nn.Cell):
84        def __init__(self, strategy1, strategy2, strategy3):
85            super().__init__()
86            self.conv1 = NetConv(16, 8, (3, 3), bias_init='zeros', strategy=strategy1)
87            self.mul1 = P.Mul().shard(strategy2)
88            self.conv2 = NetConv(8, 64, (9, 9), bias_init='zeros', strategy=strategy1)
89            self.mul2 = P.Mul().shard(strategy3)
90
91        def construct(self, x, w1, w2):
92            out1 = self.conv1(x)
93            out2 = self.mul1(out1, w1)
94            out3 = self.conv2(out2)
95            out4 = self.mul2(out3, w2)
96
97            return out4
98
99    context.set_auto_parallel_context(device_num=8, global_rank=0)
100    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
101    strategy1 = ((8, 1, 1, 1), (1, 1, 1, 1))
102    strategy2 = ((1, 1, 1, 8), (1, 1, 1, 8))
103    strategy3 = ((4, 1, 1, 2), (4, 1, 1, 2))
104
105    net = GradWrap(NetWithLoss(Net(strategy1, strategy2, strategy3)))
106    net.set_auto_parallel()
107
108    x = Tensor(np.ones([128, 16, 34, 34]), dtype=ms.float32)
109    w1 = Tensor(np.ones([128, 8, 32, 32]), dtype=ms.float32)
110    w2 = Tensor(np.ones([128, 64, 24, 24]), dtype=ms.float32)
111    net.set_train()
112    _cell_graph_executor.compile(net, x, w1, w2)
113
114
115if __name__ == '__main__':
116    test_batch()
117