• 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# ============================================================================
15import os
16import shutil
17
18import sys
19
20from tests.security_utils import security_off_wrap
21import pytest
22
23from mindspore import dataset as ds
24from mindspore import nn, Tensor, context
25from mindspore.nn.metrics import Accuracy
26from mindspore.nn.optim import Momentum
27from mindspore.dataset.transforms import c_transforms as C
28from mindspore.dataset.vision import c_transforms as CV
29from mindspore.dataset.vision import Inter
30from mindspore.common import dtype as mstype
31from mindspore.common.initializer import TruncatedNormal
32from mindspore.train import Model
33from mindspore.profiler import Profiler
34
35
36def conv(in_channels, out_channels, kernel_size, stride=1, padding=0):
37    """weight initial for conv layer"""
38    weight = weight_variable()
39    return nn.Conv2d(in_channels, out_channels,
40                     kernel_size=kernel_size, stride=stride, padding=padding,
41                     weight_init=weight, has_bias=False, pad_mode="valid")
42
43
44def fc_with_initialize(input_channels, out_channels):
45    """weight initial for fc layer"""
46    weight = weight_variable()
47    bias = weight_variable()
48    return nn.Dense(input_channels, out_channels, weight, bias)
49
50
51def weight_variable():
52    """weight initial"""
53    return TruncatedNormal(0.02)
54
55
56class LeNet5(nn.Cell):
57    """Define LeNet5 network."""
58
59    def __init__(self, num_class=10, channel=1):
60        super(LeNet5, self).__init__()
61        self.num_class = num_class
62        self.conv1 = conv(channel, 6, 5)
63        self.conv2 = conv(6, 16, 5)
64        self.fc1 = fc_with_initialize(16 * 5 * 5, 120)
65        self.fc2 = fc_with_initialize(120, 84)
66        self.fc3 = fc_with_initialize(84, self.num_class)
67        self.relu = nn.ReLU()
68        self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2)
69        self.flatten = nn.Flatten()
70        self.channel = Tensor(channel)
71
72    def construct(self, data):
73        """define construct."""
74        output = self.conv1(data)
75        output = self.relu(output)
76        output = self.max_pool2d(output)
77        output = self.conv2(output)
78        output = self.relu(output)
79        output = self.max_pool2d(output)
80        output = self.flatten(output)
81        output = self.fc1(output)
82        output = self.relu(output)
83        output = self.fc2(output)
84        output = self.relu(output)
85        output = self.fc3(output)
86        return output
87
88
89def create_dataset(data_path, batch_size=32, repeat_size=1, num_parallel_workers=1):
90    """create dataset for train"""
91    # define dataset
92    mnist_ds = ds.MnistDataset(data_path, num_samples=batch_size * 100)
93
94    resize_height, resize_width = 32, 32
95    rescale = 1.0 / 255.0
96    rescale_nml = 1 / 0.3081
97    shift_nml = -1 * 0.1307 / 0.3081
98
99    # define map operations
100    resize_op = CV.Resize((resize_height, resize_width), interpolation=Inter.LINEAR)  # Bilinear mode
101    rescale_nml_op = CV.Rescale(rescale_nml, shift_nml)
102    rescale_op = CV.Rescale(rescale, shift=0.0)
103    hwc2chw_op = CV.HWC2CHW()
104    type_cast_op = C.TypeCast(mstype.int32)
105
106    # apply map operations on images
107    mnist_ds = mnist_ds.map(operations=type_cast_op, input_columns="label", num_parallel_workers=num_parallel_workers)
108    mnist_ds = mnist_ds.map(operations=resize_op, input_columns="image", num_parallel_workers=num_parallel_workers)
109    mnist_ds = mnist_ds.map(operations=rescale_op, input_columns="image", num_parallel_workers=num_parallel_workers)
110    mnist_ds = mnist_ds.map(operations=rescale_nml_op, input_columns="image", num_parallel_workers=num_parallel_workers)
111    mnist_ds = mnist_ds.map(operations=hwc2chw_op, input_columns="image", num_parallel_workers=num_parallel_workers)
112
113    # apply DatasetOps
114    mnist_ds = mnist_ds.batch(batch_size, drop_remainder=True)
115    mnist_ds = mnist_ds.repeat(repeat_size)
116
117    return mnist_ds
118
119
120def cleanup():
121    data_path = os.path.join(os.getcwd(), "data")
122    kernel_meta_path = os.path.join(os.getcwd(), "kernel_data")
123    cache_path = os.path.join(os.getcwd(), "__pycache__")
124    if os.path.exists(data_path):
125        shutil.rmtree(data_path)
126    if os.path.exists(kernel_meta_path):
127        shutil.rmtree(kernel_meta_path)
128    if os.path.exists(cache_path):
129        shutil.rmtree(cache_path)
130
131
132class TestProfiler:
133    device_id = int(os.getenv('DEVICE_ID')) if os.getenv('DEVICE_ID') else 0
134    rank_id = int(os.getenv('RANK_ID')) if os.getenv('RANK_ID') else 0
135    mnist_path = '/home/workspace/mindspore_dataset/mnist'
136
137    @classmethod
138    def setup_class(cls):
139        """Run begin all test case start."""
140        cleanup()
141
142    @staticmethod
143    def teardown():
144        """Run after each test case end."""
145        cleanup()
146
147    @pytest.mark.level2
148    @pytest.mark.platform_x86_cpu
149    @pytest.mark.env_onecard
150    @security_off_wrap
151    def test_cpu_profiler(self):
152        if sys.platform != 'linux':
153            return
154        self._train_with_profiler(device_target="CPU")
155        self._check_cpu_profiling_file()
156
157    @pytest.mark.level1
158    @pytest.mark.platform_x86_gpu_training
159    @pytest.mark.env_onecard
160    @security_off_wrap
161    def test_gpu_profiler(self):
162        self._train_with_profiler(device_target="GPU")
163        self._check_gpu_profiling_file()
164
165    @pytest.mark.level0
166    @pytest.mark.platform_arm_ascend_training
167    @pytest.mark.platform_x86_ascend_training
168    @pytest.mark.env_onecard
169    @security_off_wrap
170    def test_ascend_profiler(self):
171        self._train_with_profiler(device_target="Ascend")
172        self._check_d_profiling_file()
173
174    def _train_with_profiler(self, device_target):
175        context.set_context(mode=context.GRAPH_MODE, device_target=device_target)
176        profiler = Profiler(profile_memory=True, output_path='data')
177        profiler_name = os.listdir(os.path.join(os.getcwd(), 'data'))[0]
178        self.profiler_path = os.path.join(os.getcwd(), f'data/{profiler_name}/')
179        ds_train = create_dataset(os.path.join(self.mnist_path, "train"))
180        if ds_train.get_dataset_size() == 0:
181            raise ValueError("Please check dataset size > 0 and batch_size <= dataset size")
182
183        lenet = LeNet5()
184        loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction="mean")
185        optim = Momentum(lenet.trainable_params(), learning_rate=0.1, momentum=0.9)
186        model = Model(lenet, loss_fn=loss, optimizer=optim, metrics={'acc': Accuracy()})
187
188        model.train(1, ds_train, dataset_sink_mode=True)
189        profiler.analyse()
190
191    def _check_gpu_profiling_file(self):
192        op_detail_file = self.profiler_path + f'gpu_op_detail_info_{self.device_id}.csv'
193        op_type_file = self.profiler_path + f'gpu_op_type_info_{self.device_id}.csv'
194        activity_file = self.profiler_path + f'gpu_activity_data_{self.device_id}.csv'
195        timeline_file = self.profiler_path + f'gpu_timeline_display_{self.device_id}.json'
196        getnext_file = self.profiler_path + f'minddata_getnext_profiling_{self.device_id}.txt'
197        pipeline_file = self.profiler_path + f'minddata_pipeline_raw_{self.device_id}.csv'
198
199        gpu_profiler_files = (op_detail_file, op_type_file, activity_file,
200                              timeline_file, getnext_file, pipeline_file)
201        for file in gpu_profiler_files:
202            assert os.path.isfile(file)
203
204    def _check_d_profiling_file(self):
205        aicore_file = self.profiler_path + f'aicore_intermediate_{self.rank_id}_detail.csv'
206        step_trace_file = self.profiler_path + f'step_trace_raw_{self.rank_id}_detail_time.csv'
207        timeline_file = self.profiler_path + f'ascend_timeline_display_{self.rank_id}.json'
208        aicpu_file = self.profiler_path + f'aicpu_intermediate_{self.rank_id}.csv'
209        minddata_pipeline_file = self.profiler_path + f'minddata_pipeline_raw_{self.rank_id}.csv'
210        queue_profiling_file = self.profiler_path + f'device_queue_profiling_{self.rank_id}.txt'
211        memory_file = self.profiler_path + f'memory_usage_{self.rank_id}.pb'
212
213        d_profiler_files = (aicore_file, step_trace_file, timeline_file, aicpu_file,
214                            minddata_pipeline_file, queue_profiling_file, memory_file)
215        for file in d_profiler_files:
216            assert os.path.isfile(file)
217
218    def _check_cpu_profiling_file(self):
219        op_detail_file = self.profiler_path + f'cpu_op_detail_info_{self.device_id}.csv'
220        op_type_file = self.profiler_path + f'cpu_op_type_info_{self.device_id}.csv'
221        timeline_file = self.profiler_path + f'cpu_op_execute_timestamp_{self.device_id}.txt'
222
223        cpu_profiler_files = (op_detail_file, op_type_file, timeline_file)
224        for file in cpu_profiler_files:
225            assert os.path.isfile(file)
226