• 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
16"""Bert test."""
17
18# pylint: disable=missing-docstring, arguments-differ, W0612
19
20import os
21
22import mindspore.common.dtype as mstype
23import mindspore.context as context
24from mindspore import Tensor
25from mindspore.ops import operations as P
26from mindspore.nn.optim import AdamWeightDecay
27from mindspore.train.loss_scale_manager import DynamicLossScaleManager
28from mindspore.nn import learning_rate_schedule as lr_schedules
29from tests.models.official.nlp.bert.src import BertConfig, BertNetworkWithLoss, BertTrainOneStepCell, BertTrainOneStepWithLossScaleCell
30from ...dataset_mock import MindData
31from ...ops_common import nn, np, batch_tuple_tensor, build_construct_graph
32
33
34_current_dir = os.path.dirname(os.path.realpath(__file__)) + "/../python/test_data"
35context.set_context(mode=context.GRAPH_MODE)
36
37
38def get_dataset(batch_size=1):
39    dataset_types = (np.int32, np.int32, np.int32, np.int32, np.int32, np.int32, np.int32)
40    dataset_shapes = ((batch_size, 128), (batch_size, 128), (batch_size, 128), (batch_size, 1), \
41                      (batch_size, 20), (batch_size, 20), (batch_size, 20))
42
43    dataset = MindData(size=2, batch_size=batch_size,
44                       np_types=dataset_types,
45                       output_shapes=dataset_shapes,
46                       input_indexs=(0, 1))
47    return dataset
48
49
50def load_test_data(batch_size=1):
51    dataset = get_dataset(batch_size)
52    ret = dataset.next()
53    ret = batch_tuple_tensor(ret, batch_size)
54    return ret
55
56
57def get_config(version='base'):
58    """
59    get_config definition
60    """
61    if version == 'base':
62        return BertConfig(
63            seq_length=128,
64            vocab_size=21128,
65            hidden_size=768,
66            num_hidden_layers=12,
67            num_attention_heads=12,
68            intermediate_size=3072,
69            hidden_act="gelu",
70            hidden_dropout_prob=0.1,
71            attention_probs_dropout_prob=0.1,
72            max_position_embeddings=512,
73            type_vocab_size=2,
74            initializer_range=0.02,
75            use_relative_positions=True,
76            dtype=mstype.float32,
77            compute_type=mstype.float32)
78    if version == 'large':
79        return BertConfig(
80            seq_length=128,
81            vocab_size=21128,
82            hidden_size=1024,
83            num_hidden_layers=24,
84            num_attention_heads=16,
85            intermediate_size=4096,
86            hidden_act="gelu",
87            hidden_dropout_prob=0.1,
88            attention_probs_dropout_prob=0.1,
89            max_position_embeddings=512,
90            type_vocab_size=2,
91            initializer_range=0.02,
92            use_relative_positions=True,
93            dtype=mstype.float32,
94            compute_type=mstype.float32)
95    return BertConfig()
96
97
98class BertLearningRate(lr_schedules.LearningRateSchedule):
99    def __init__(self, decay_steps, warmup_steps=100, learning_rate=0.1, end_learning_rate=0.0001, power=1.0):
100        super(BertLearningRate, self).__init__()
101        self.warmup_lr = lr_schedules.WarmUpLR(learning_rate, warmup_steps)
102        self.decay_lr = lr_schedules.PolynomialDecayLR(learning_rate, end_learning_rate, decay_steps, power)
103        self.warmup_steps = Tensor(np.array([warmup_steps]).astype(np.float32))
104
105        self.greater = P.Greater()
106        self.one = Tensor(np.array([1.0]).astype(np.float32))
107        self.cast = P.Cast()
108
109    def construct(self, global_step):
110        is_warmup = self.cast(self.greater(self.warmup_steps, global_step), mstype.float32)
111        warmup_lr = self.warmup_lr(global_step)
112        decay_lr = self.decay_lr(global_step)
113        lr = (self.one - is_warmup) * decay_lr + is_warmup * warmup_lr
114        return lr
115
116
117def test_bert_train():
118    """
119    the main function
120    """
121
122    class ModelBert(nn.Cell):
123        """
124        ModelBert definition
125        """
126
127        def __init__(self, network, optimizer=None):
128            super(ModelBert, self).__init__()
129            self.optimizer = optimizer
130            self.train_network = BertTrainOneStepCell(network, self.optimizer)
131            self.train_network.set_train()
132
133        def construct(self, arg0, arg1, arg2, arg3, arg4, arg5, arg6):
134            return self.train_network(arg0, arg1, arg2, arg3, arg4, arg5, arg6)
135
136    version = os.getenv('VERSION', 'large')
137    batch_size = int(os.getenv('BATCH_SIZE', '1'))
138    inputs = load_test_data(batch_size)
139
140    config = get_config(version=version)
141    netwithloss = BertNetworkWithLoss(config, True)
142    lr = BertLearningRate(10)
143    optimizer = AdamWeightDecay(netwithloss.trainable_params(), lr)
144    net = ModelBert(netwithloss, optimizer=optimizer)
145    net.set_train()
146    build_construct_graph(net, *inputs, execute=False)
147
148
149def test_bert_withlossscale_train():
150    class ModelBert(nn.Cell):
151        def __init__(self, network, optimizer=None):
152            super(ModelBert, self).__init__()
153            self.optimizer = optimizer
154            self.train_network = BertTrainOneStepWithLossScaleCell(network, self.optimizer)
155            self.train_network.set_train()
156
157        def construct(self, arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7):
158            return self.train_network(arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7)
159
160    version = os.getenv('VERSION', 'base')
161    batch_size = int(os.getenv('BATCH_SIZE', '1'))
162    scaling_sens = Tensor(np.ones([1]).astype(np.float32))
163    inputs = load_test_data(batch_size) + (scaling_sens,)
164
165    config = get_config(version=version)
166    netwithloss = BertNetworkWithLoss(config, True)
167    lr = BertLearningRate(10)
168    optimizer = AdamWeightDecay(netwithloss.trainable_params(), lr)
169    net = ModelBert(netwithloss, optimizer=optimizer)
170    net.set_train()
171    build_construct_graph(net, *inputs, execute=True)
172
173
174def bert_withlossscale_manager_train():
175    class ModelBert(nn.Cell):
176        def __init__(self, network, optimizer=None):
177            super(ModelBert, self).__init__()
178            self.optimizer = optimizer
179            manager = DynamicLossScaleManager()
180            update_cell = LossScaleUpdateCell(manager)
181            self.train_network = BertTrainOneStepWithLossScaleCell(network, self.optimizer,
182                                                                   scale_update_cell=update_cell)
183            self.train_network.set_train()
184
185        def construct(self, arg0, arg1, arg2, arg3, arg4, arg5, arg6):
186            return self.train_network(arg0, arg1, arg2, arg3, arg4, arg5, arg6)
187
188    version = os.getenv('VERSION', 'base')
189    batch_size = int(os.getenv('BATCH_SIZE', '1'))
190    inputs = load_test_data(batch_size)
191
192    config = get_config(version=version)
193    netwithloss = BertNetworkWithLoss(config, True)
194    lr = BertLearningRate(10)
195    optimizer = AdamWeightDecay(netwithloss.trainable_params(), lr)
196    net = ModelBert(netwithloss, optimizer=optimizer)
197    net.set_train()
198    build_construct_graph(net, *inputs, execute=True)
199
200
201def bert_withlossscale_manager_train_feed():
202    class ModelBert(nn.Cell):
203        def __init__(self, network, optimizer=None):
204            super(ModelBert, self).__init__()
205            self.optimizer = optimizer
206            manager = DynamicLossScaleManager()
207            update_cell = LossScaleUpdateCell(manager)
208            self.train_network = BertTrainOneStepWithLossScaleCell(network, self.optimizer,
209                                                                   scale_update_cell=update_cell)
210            self.train_network.set_train()
211
212        def construct(self, arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7):
213            return self.train_network(arg0, arg1, arg2, arg3, arg4, arg5, arg6, arg7)
214
215    version = os.getenv('VERSION', 'base')
216    batch_size = int(os.getenv('BATCH_SIZE', '1'))
217    scaling_sens = Tensor(np.ones([1]).astype(np.float32))
218    inputs = load_test_data(batch_size) + (scaling_sens,)
219
220    config = get_config(version=version)
221    netwithloss = BertNetworkWithLoss(config, True)
222    lr = BertLearningRate(10)
223    optimizer = AdamWeightDecay(netwithloss.trainable_params(), lr)
224    net = ModelBert(netwithloss, optimizer=optimizer)
225    net.set_train()
226    build_construct_graph(net, *inputs, execute=True)
227