• 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# ============================================================================
15from collections import Counter
16import numpy as np
17
18import mindspore.nn as nn
19from mindspore import Tensor, Parameter
20from mindspore.common import dtype as mstype
21from mindspore.common.api import _cell_graph_executor
22from mindspore.nn import TrainOneStepCell, WithLossCell
23from mindspore.nn.optim import LARS, Momentum
24from mindspore.ops import operations as P
25
26
27def multisteplr(total_steps, milestone, base_lr=0.9, gamma=0.1, dtype=mstype.float32):
28    lr = []
29    milestone = Counter(milestone)
30
31    for step in range(total_steps):
32        base_lr = base_lr * gamma ** milestone[step]
33        lr.append(base_lr)
34    return Tensor(np.array(lr), dtype)
35
36
37class Net(nn.Cell):
38    def __init__(self):
39        super(Net, self).__init__()
40        self.weight = Parameter(Tensor(np.ones([64, 10]).astype(np.float32)), name="weight")
41        self.bias = Parameter(Tensor(np.ones([10]).astype((np.float32))), name="bias")
42        self.matmul = P.MatMul()
43        self.biasAdd = P.BiasAdd()
44
45    def construct(self, x):
46        x = self.biasAdd(self.matmul(x, self.weight), self.bias)
47        return x
48
49
50def test_lars_multi_step_lr():
51    inputs = Tensor(np.ones([1, 64]).astype(np.float32))
52    label = Tensor(np.zeros([1, 10]).astype(np.float32))
53    net = Net()
54    net.set_train()
55    loss = nn.SoftmaxCrossEntropyWithLogits()
56
57    lr = multisteplr(10, [2, 6])
58    SGD = Momentum(net.trainable_params(), lr, 0.9)
59    optimizer = LARS(SGD, epsilon=1e-08, coefficient=0.02, use_clip=True,
60                     lars_filter=lambda x: 'bn' not in x.name)
61
62    net_with_loss = WithLossCell(net, loss)
63    train_network = TrainOneStepCell(net_with_loss, optimizer)
64    _cell_graph_executor.compile(train_network, inputs, label)
65
66
67def test_lars_float_lr():
68    inputs = Tensor(np.ones([1, 64]).astype(np.float32))
69    label = Tensor(np.zeros([1, 10]).astype(np.float32))
70    net = Net()
71    net.set_train()
72    loss = nn.SoftmaxCrossEntropyWithLogits()
73
74    lr = 0.1
75    SGD = Momentum(net.trainable_params(), lr, 0.9)
76    optimizer = LARS(SGD, epsilon=1e-08, coefficient=0.02,
77                     lars_filter=lambda x: 'bn' not in x.name)
78
79    net_with_loss = WithLossCell(net, loss)
80    train_network = TrainOneStepCell(net_with_loss, optimizer)
81    _cell_graph_executor.compile(train_network, inputs, label)
82