• 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
18
19import mindspore.context as context
20import mindspore.nn as nn
21from mindspore import Tensor
22from mindspore.common.api import ms_function
23from mindspore.ops import operations as P
24from mindspore.ops import functional as F
25from mindspore.common import dtype as mstype
26from mindspore.common.parameter import Parameter
27
28
29class Net(nn.Cell):
30    def __init__(self, decay_flag=True):
31        super(Net, self).__init__()
32        self.decay_flag = decay_flag
33        self.op_mul = P.Mul()
34        self.op_square = P.Square()
35        self.op_sqrt = P.Sqrt()
36        self.op_cast = P.Cast()
37        self.op_reshape = P.Reshape()
38        self.op_shape = P.Shape()
39        self.param = Parameter(
40            Tensor(np.array([1, 3, 5]).astype(np.float32)), name='param')
41        self.m = Parameter(
42            Tensor(np.array([0.11, 0.33, 0.55]).astype(np.float32)), name='m')
43        self.v = Parameter(
44            Tensor(np.array([1.2, 3.4, 5.6]).astype(np.float32)), name='v')
45
46    @ms_function
47    def construct(self, beta1, beta2, one_sub_beta_1, one_sub_beta_2, gradient, eps, weight_decay_tensor, lr):
48        param_fp32 = self.op_cast(self.param, mstype.float32)
49        m_fp32 = self.op_cast(self.m, mstype.float32)
50        v_fp32 = self.op_cast(self.v, mstype.float32)
51        gradient_fp32 = self.op_cast(gradient, mstype.float32)
52
53        next_m = self.op_mul(beta1, m_fp32) + \
54            self.op_mul(self.op_cast(one_sub_beta_1,
55                                     mstype.float32), gradient_fp32)
56        next_v = self.op_mul(beta2, v_fp32) + self.op_mul(self.op_cast(one_sub_beta_2,
57                                                                       mstype.float32), self.op_square(gradient_fp32))
58        update = next_m / (eps + self.op_sqrt(next_v))
59        if self.decay_flag:
60            update = self.op_mul(weight_decay_tensor, param_fp32) + update
61        update_with_lr = self.op_mul(lr, update)
62        next_param = param_fp32 - \
63            self.op_reshape(update_with_lr, self.op_shape(param_fp32))
64
65        depend_v = F.depend(next_param, F.assign(self.param, next_param))
66        depend_v = F.depend(depend_v, F.assign(self.m, next_m))
67        depend_v = F.depend(depend_v, F.assign(self.v, next_v))
68        return depend_v
69
70
71class SideEffectFusedAdamNet(nn.Cell):
72    def __init__(self, decay_flag=True):
73        super(SideEffectFusedAdamNet, self).__init__()
74        self.decay_flag = decay_flag
75        self.op_mul = P.Mul()
76        self.op_square = P.Square()
77        self.op_sqrt = P.Sqrt()
78        self.op_cast = P.Cast()
79        self.op_reshape = P.Reshape()
80        self.op_shape = P.Shape()
81        self.param = Parameter(
82            Tensor(np.array([0, 0, 0]).astype(np.float32)), name='param')
83        self.m = Parameter(
84            Tensor(np.array([0.11, 0.33, 0.55]).astype(np.float32)), name='m')
85        self.v = Parameter(
86            Tensor(np.array([1.2, 3.4, 5.6]).astype(np.float32)), name='v')
87        self.x = Parameter(
88            Tensor(np.array([1, 3, 5]).astype(np.float32)), name='x')
89
90    @ms_function
91    def construct(self, beta1, beta2, one_sub_beta_1, one_sub_beta_2, gradient, eps, weight_decay_tensor, lr):
92        F.assign(self.param, self.x)
93
94        param_fp32 = self.op_cast(self.param, mstype.float32)
95        m_fp32 = self.op_cast(self.m, mstype.float32)
96        v_fp32 = self.op_cast(self.v, mstype.float32)
97        gradient_fp32 = self.op_cast(gradient, mstype.float32)
98
99        next_m = self.op_mul(beta1, m_fp32) + \
100            self.op_mul(self.op_cast(one_sub_beta_1,
101                                     mstype.float32), gradient_fp32)
102        next_v = self.op_mul(beta2, v_fp32) + self.op_mul(self.op_cast(one_sub_beta_2,
103                                                                       mstype.float32), self.op_square(gradient_fp32))
104        update = next_m / (eps + self.op_sqrt(next_v))
105        if self.decay_flag:
106            update = self.op_mul(weight_decay_tensor, param_fp32) + update
107        update_with_lr = self.op_mul(lr, update)
108        next_param = param_fp32 - \
109            self.op_reshape(update_with_lr, self.op_shape(param_fp32))
110
111        depend_v = F.depend(next_param, F.assign(self.param, next_param))
112        depend_v = F.depend(depend_v, F.assign(self.m, next_m))
113        depend_v = F.depend(depend_v, F.assign(self.v, next_v))
114
115        F.assign(self.x, self.m)
116        return depend_v
117
118
119def CalFusedAdam(beta1, beta2, one_sub_beta_1, one_sub_beta_2, gradient, eps, weight_decay_tensor, lr, param, m, v,
120                 is_weight_decay=False):
121    m_expect = beta1 * m + one_sub_beta_1 * gradient
122    v_expect = beta2 * v + one_sub_beta_2 * gradient * gradient
123    update = m_expect / (np.sqrt(v_expect) + eps)
124    if is_weight_decay:
125        update += weight_decay_tensor * param
126    param_expect = param - lr * update
127    return param_expect, m_expect, v_expect
128
129
130def test_adam():
131    np.random.seed(0)
132    beta1 = np.array([0.9]).astype(np.float32)
133    beta2 = np.array([0.999]).astype(np.float32)
134    one_sub_beta_1 = (np.array([1.0]) - np.array([0.9])).astype(np.float32)
135    one_sub_beta_2 = (np.array([1.0]) - np.array([0.999])).astype(np.float32)
136    lr = np.array([0.012]).astype(np.float32)
137    eps = np.array([1e-6]).astype(np.float32)
138    weight_decay_tensor = np.array([0.021]).astype(np.float32)
139
140    gradient = np.array([0.01, 0.03, 0.05]).astype(np.float32)
141    m = np.array([0.11, 0.33, 0.55]).astype(np.float32)
142    v = np.array([1.2, 3.4, 5.6]).astype(np.float32)
143    param = np.array([1, 3, 5]).astype(np.float32)
144    is_weight_decay = False
145    opt = Net(is_weight_decay)
146    _ = opt(Tensor(beta1), Tensor(beta2), Tensor(one_sub_beta_1), Tensor(one_sub_beta_2), Tensor(gradient), Tensor(eps),
147            Tensor(weight_decay_tensor), Tensor(lr))
148    param_expect, m_expect, v_expect = CalFusedAdam(
149        beta1, beta2, one_sub_beta_1, one_sub_beta_2, gradient, eps, weight_decay_tensor, lr,
150        param, m, v, is_weight_decay)
151    assert np.allclose(opt.param.data.asnumpy(), param_expect,
152                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
153    assert np.allclose(opt.m.data.asnumpy(), m_expect,
154                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
155    assert np.allclose(opt.v.data.asnumpy(), v_expect,
156                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
157
158
159def test_adam_weight_decay():
160    np.random.seed(0)
161    beta1 = np.array([0.9]).astype(np.float32)
162    beta2 = np.array([0.999]).astype(np.float32)
163    one_sub_beta_1 = (np.array([1.0]) - np.array([0.9])).astype(np.float32)
164    one_sub_beta_2 = (np.array([1.0]) - np.array([0.999])).astype(np.float32)
165    lr = np.array([0.012]).astype(np.float32)
166    eps = np.array([1e-6]).astype(np.float32)
167    weight_decay_tensor = np.array([0.021]).astype(np.float32)
168
169    gradient = np.array([0.01, 0.03, 0.05]).astype(np.float32)
170    m = np.array([0.11, 0.33, 0.55]).astype(np.float32)
171    v = np.array([1.2, 3.4, 5.6]).astype(np.float32)
172    param = np.array([1, 3, 5]).astype(np.float32)
173    is_weight_decay = True
174    opt = Net(is_weight_decay)
175    _ = opt(Tensor(beta1), Tensor(beta2), Tensor(one_sub_beta_1), Tensor(one_sub_beta_2), Tensor(gradient), Tensor(eps),
176            Tensor(weight_decay_tensor), Tensor(lr))
177    param_expect, m_expect, v_expect = CalFusedAdam(
178        beta1, beta2, one_sub_beta_1, one_sub_beta_2, gradient, eps, weight_decay_tensor, lr,
179        param, m, v, is_weight_decay)
180
181    assert np.allclose(opt.param.data.asnumpy(), param_expect,
182                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
183    assert np.allclose(opt.m.data.asnumpy(), m_expect,
184                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
185    assert np.allclose(opt.v.data.asnumpy(), v_expect,
186                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
187
188
189def test_adam_side_effect():
190    np.random.seed(0)
191    beta1 = np.array([0.9]).astype(np.float32)
192    beta2 = np.array([0.999]).astype(np.float32)
193    one_sub_beta_1 = (np.array([1.0]) - np.array([0.9])).astype(np.float32)
194    one_sub_beta_2 = (np.array([1.0]) - np.array([0.999])).astype(np.float32)
195    lr = np.array([0.012]).astype(np.float32)
196    eps = np.array([1e-6]).astype(np.float32)
197    weight_decay_tensor = np.array([0.021]).astype(np.float32)
198
199    gradient = np.array([0.01, 0.03, 0.05]).astype(np.float32)
200    m = np.array([0.11, 0.33, 0.55]).astype(np.float32)
201    v = np.array([1.2, 3.4, 5.6]).astype(np.float32)
202    param = np.array([1, 3, 5]).astype(np.float32)
203    is_weight_decay = False
204    opt = SideEffectFusedAdamNet(is_weight_decay)
205    _ = opt(Tensor(beta1), Tensor(beta2), Tensor(one_sub_beta_1), Tensor(one_sub_beta_2), Tensor(gradient), Tensor(eps),
206            Tensor(weight_decay_tensor), Tensor(lr))
207    param_expect, m_expect, v_expect = CalFusedAdam(
208        beta1, beta2, one_sub_beta_1, one_sub_beta_2, gradient, eps, weight_decay_tensor, lr,
209        param, m, v, is_weight_decay)
210    assert np.allclose(opt.param.data.asnumpy(), param_expect,
211                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
212    assert np.allclose(opt.m.data.asnumpy(), m_expect,
213                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
214    assert np.allclose(opt.v.data.asnumpy(), v_expect,
215                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
216    assert np.allclose(opt.x.data.asnumpy(), m_expect,
217                       rtol=1.e-4, atol=1.e-8, equal_nan=True)
218
219
220@pytest.mark.level0
221@pytest.mark.platform_x86_gpu_training
222@pytest.mark.env_onecard
223def test_adam_gpu():
224    context.set_context(mode=context.GRAPH_MODE,
225                        enable_graph_kernel=True, device_target="GPU")
226    test_adam()
227
228
229def test_adam_ascend():
230    context.set_context(mode=context.GRAPH_MODE,
231                        enable_graph_kernel=True, device_target="Ascend")
232    test_adam()
233
234
235@pytest.mark.level0
236@pytest.mark.platform_x86_gpu_training
237@pytest.mark.env_onecard
238def test_adam_weight_decay_gpu():
239    context.set_context(mode=context.GRAPH_MODE,
240                        enable_graph_kernel=True, device_target="GPU")
241    test_adam_weight_decay()
242
243
244def test_adam_weight_decay_ascend():
245    context.set_context(mode=context.GRAPH_MODE,
246                        enable_graph_kernel=True, device_target="Ascend")
247    test_adam_weight_decay()
248
249
250@pytest.mark.level0
251@pytest.mark.platform_x86_gpu_training
252@pytest.mark.env_onecard
253def test_adam_side_effect_gpu():
254    context.set_context(mode=context.GRAPH_MODE,
255                        enable_graph_kernel=True, device_target="GPU")
256    test_adam_side_effect()
257
258
259@pytest.mark.level2
260@pytest.mark.platform_arm_ascend_training
261@pytest.mark.platform_x86_ascend_training
262@pytest.mark.env_onecard
263def test_adam_side_effect_ascend():
264    context.set_context(mode=context.GRAPH_MODE,
265                        enable_graph_kernel=True, device_target="Ascend")
266    test_adam_side_effect()
267