• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2020 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# ============================================================================
15
16import numpy as np
17import pytest
18import mindspore as ms
19from mindspore import context, Tensor, Parameter
20from mindspore.common.api import _cell_graph_executor
21from mindspore.nn import Cell, TrainOneStepCell, Momentum
22from mindspore.ops import operations as P
23from mindspore.common.initializer import initializer
24
25class Net(Cell):
26    def __init__(self,
27                 strategy1=None,
28                 strategy2=None,
29                 strategy3=None,
30                 axis=0,
31                 init_flag=True,
32                 split_tuple=(4, 4),
33                 split_string="manual_split",
34                 param_shape=(8, 8)):
35        super().__init__()
36        self.gatherv2 = P.Gather().shard(strategy1)
37        self.gatherv2.add_prim_attr(split_string, split_tuple)
38        self.mul = P.Mul().shard(strategy2)
39        self.reshape = P.Reshape()
40        self.matmul = P.MatMul().shard(strategy3)
41        self.matmul.add_prim_attr("forward_reduce_scatter", True)
42        if init_flag:
43            self.param = Parameter(initializer("ones", param_shape, ms.float32), name="gatherv2_param")
44        else:
45            self.param = Parameter(Tensor(np.ones(param_shape), dtype=ms.float32), name="gatherv2_param")
46        self.mul_weight = Parameter(initializer("ones", (8, 8, 8), ms.float32), name="mul_weight")
47        self.matmul_weight = Parameter(initializer("ones", (64, 16), ms.float32), name="matmul_weight")
48        self.axis = axis
49
50    def construct(self, x, b):
51        out = self.gatherv2(self.param, x, self.axis)
52        out = self.mul(out, self.mul_weight)
53        out = self.reshape(out, (8, 64))
54        out = self.matmul(out, self.matmul_weight)
55        return out
56
57
58_x = Tensor(np.ones([8, 8]), dtype=ms.int32)
59_b = Tensor(np.ones([64, 8]), dtype=ms.float32)
60
61
62def compile_net(net):
63    optimizer = Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
64    train_net = TrainOneStepCell(net, optimizer)
65    train_net.set_auto_parallel()
66    train_net.set_train()
67    _cell_graph_executor.compile(train_net, _x, _b, auto_parallel_mode=True)
68    context.reset_auto_parallel_context()
69
70
71def test_normal_split():
72    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
73    strategy1 = ((2, 1), (1, 2))
74    strategy2 = ((1, 2, 1), (1, 2, 1))
75    strategy3 = ((1, 2), (2, 1))
76    net = Net(strategy1, strategy2, strategy3)
77    compile_net(net)
78
79
80def test_normal_split2():
81    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=4, global_rank=0)
82    strategy1 = ((4, 1), (1, 4))
83    strategy2 = ((1, 4, 1), (1, 4, 1))
84    strategy3 = ((1, 4), (4, 1))
85    net = Net(strategy1, strategy2, strategy3, split_tuple=(10, 20, 30, 4), param_shape=(64, 8))
86    compile_net(net)
87
88
89def test_normal_split3():
90    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=32, global_rank=17)
91    strategy1 = ((4, 8), (1, 4))
92    strategy2 = ((1, 4, 8), (1, 4, 8))
93    strategy3 = ((1, 32), (32, 1))
94    net = Net(strategy1, strategy2, strategy3, split_tuple=(10, 20, 30, 4), param_shape=(64, 8))
95    compile_net(net)
96
97
98def test_normal_split_with_offset():
99    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
100    strategy1 = ((2, 1), (1, 2))
101    strategy2 = ((1, 2, 1), (1, 2, 1))
102    strategy3 = ((1, 2), (2, 1))
103    net = Net(strategy1, strategy2, strategy3, split_string="manual_split_with_offset", split_tuple=((4, 0), (4, 4)))
104    compile_net(net)
105
106
107def test_auto_parallel_error():
108    context.set_auto_parallel_context(parallel_mode="auto_parallel", device_num=2, global_rank=0)
109    net = Net()
110    with pytest.raises(RuntimeError):
111        compile_net(net)
112
113
114def test_axis_error():
115    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
116    strategy1 = ((2, 1), (1, 2))
117    strategy2 = ((1, 2, 1), (1, 2, 1))
118    strategy3 = ((1, 2), (2, 1))
119    net = Net(strategy1, strategy2, strategy3, axis=1)
120    with pytest.raises(RuntimeError):
121        compile_net(net)
122
123
124def test_strategy_error():
125    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
126    strategy1 = ((4, 1), (8, 1))
127    strategy2 = ((1, 2, 1), (1, 2, 1))
128    strategy3 = ((1, 2), (2, 1))
129    net = Net(strategy1, strategy2, strategy3)
130    with pytest.raises(RuntimeError):
131        compile_net(net)
132
133
134def test_strategy_error2():
135    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
136    strategy1 = ((4, 1), (1, 8))
137    strategy2 = ((1, 2, 1), (1, 2, 1))
138    strategy3 = ((1, 2), (2, 1))
139    net = Net(strategy1, strategy2, strategy3)
140    with pytest.raises(RuntimeError):
141        compile_net(net)
142
143
144def test_strategy_error3():
145    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
146    strategy1 = ((2, 1), (1, 2))
147    strategy2 = ((1, 2, 1), (1, 2, 1))
148    strategy3 = ((1, 2), (2, 1))
149    net = Net(strategy1, strategy2, strategy3)
150    with pytest.raises(RuntimeError):
151        compile_net(net)
152
153
154def test_strategy_error4():
155    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
156    strategy1 = ((2, 8), (1, 2))
157    strategy2 = ((1, 2, 1), (1, 2, 1))
158    strategy3 = ((1, 2), (2, 1))
159    net = Net(strategy1, strategy2, strategy3)
160    with pytest.raises(RuntimeError):
161        compile_net(net)
162
163
164def test_strategy_error5():
165    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=4, global_rank=0)
166    strategy1 = ((4, 1), (1, 4))
167    strategy2 = ((1, 2, 1), (1, 2, 1))
168    strategy3 = ((1, 2), (2, 1))
169    net = Net(strategy1, strategy2, strategy3)
170    with pytest.raises(RuntimeError):
171        compile_net(net)
172
173
174def test_split_tuple_error():
175    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
176    strategy1 = ((2, 1), (1, 2))
177    strategy2 = ((1, 2, 1), (1, 2, 1))
178    strategy3 = ((1, 2), (2, 1))
179    net = Net(strategy1, strategy2, strategy3, split_tuple=((5, 0), (5, 5)))
180    with pytest.raises(RuntimeError):
181        compile_net(net)
182
183
184def test_parameter_use_tensor_error():
185    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=2, global_rank=0)
186    strategy1 = ((2, 1), (1, 2))
187    strategy2 = ((1, 2, 1), (1, 2, 1))
188    strategy3 = ((1, 2), (2, 1))
189    net = Net(strategy1, strategy2, strategy3, init_flag=False)
190    with pytest.raises(RuntimeError):
191        compile_net(net)
192