Huawei_Technology
/
mindspore
Not watched
1
0
Fork 0
Code
Releases 13 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
Tree: 1aa16fb431
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '1aa16fb431'
${ noResults }
mindspore/model_zoo/research/nlp/dscnn/src/config.py

162 lines
9.4 kB
Raw Blame History 

  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. """Config setting, will be used in train.py and eval.py"""

  16. from src.utils import prepare_words_list

  17. 

  18. def data_config(parser):

  19.     '''config for data.'''

  20. 

  21.     parser.add_argument('--data_url', type=str,

  22.                         default='http://download.tensorflow.org/data/speech_commands_v0.01.tar.gz',

  23.                         help='Location of speech training data archive on the web.')

  24.     parser.add_argument('--data_dir', type=str, default='data',

  25.                         help='Where to download the dataset.')

  26.     parser.add_argument('--feat_dir', type=str, default='feat',

  27.                         help='Where to save the feature of audios')

  28.     parser.add_argument('--background_volume', type=float, default=0.1,

  29.                         help='How loud the background noise should be, between 0 and 1.')

  30.     parser.add_argument('--background_frequency', type=float, default=0.8,

  31.                         help='How many of the training samples have background noise mixed in.')

  32.     parser.add_argument('--silence_percentage', type=float, default=10.0,

  33.                         help='How much of the training data should be silence.')

  34.     parser.add_argument('--unknown_percentage', type=float, default=10.0,

  35.                         help='How much of the training data should be unknown words.')

  36.     parser.add_argument('--time_shift_ms', type=float, default=100.0,

  37.                         help='Range to randomly shift the training audio by in time.')

  38.     parser.add_argument('--testing_percentage', type=int, default=10,

  39.                         help='What percentage of wavs to use as a test set.')

  40.     parser.add_argument('--validation_percentage', type=int, default=10,

  41.                         help='What percentage of wavs to use as a validation set.')

  42.     parser.add_argument('--wanted_words', type=str, default='yes,no,up,down,left,right,on,off,stop,go',

  43.                         help='Words to use (others will be added to an unknown label)')

  44.     parser.add_argument('--sample_rate', type=int, default=16000, help='Expected sample rate of the wavs')

  45.     parser.add_argument('--clip_duration_ms', type=int, default=1000,

  46.                         help='Expected duration in milliseconds of the wavs')

  47.     parser.add_argument('--window_size_ms', type=float, default=40.0, help='How long each spectrogram timeslice is')

  48.     parser.add_argument('--window_stride_ms', type=float, default=20.0, help='How long each spectrogram timeslice is')

  49.     parser.add_argument('--dct_coefficient_count', type=int, default=20,

  50.                         help='How many bins to use for the MFCC fingerprint')

  51. 

  52. 

  53. def train_config(parser):

  54.     '''config for train.'''

  55.     data_config(parser)

  56. 

  57.     # network related

  58.     parser.add_argument('--model_size_info', type=int, nargs="+",

  59.                         default=[6, 276, 10, 4, 2, 1, 276, 3, 3, 2, 2, 276, 3, 3, 1, 1, 276, 3, 3, 1, 1, 276, 3, 3, 1,

  60.                                  1, 276, 3, 3, 1, 1],

  61.                         help='Model dimensions - different for various models')

  62.     parser.add_argument('--drop', type=float, default=0.9, help='dropout')

  63.     parser.add_argument('--pretrained', type=str, default='', help='model_path, local pretrained model to load')

  64. 

  65.     # training related

  66.     parser.add_argument('--use_graph_mode', default=1, type=int, help='use graph mode or feed mode')

  67.     parser.add_argument('--val_interval', type=int, default=1, help='validate interval')

  68. 

  69.     # dataset related

  70.     parser.add_argument('--per_batch_size', default=100, type=int, help='batch size for per gpu')

  71. 

  72.     # optimizer and lr related

  73.     parser.add_argument('--lr_scheduler', default='multistep', type=str,

  74.                         help='lr-scheduler, option type: multistep, cosine_annealing')

  75.     parser.add_argument('--lr', default=0.1, type=float, help='learning rate of the training')

  76.     parser.add_argument('--lr_epochs', type=str, default='20,40,60,80', help='epoch of lr changing')

  77.     parser.add_argument('--lr_gamma', type=float, default=0.1,

  78.                         help='decrease lr by a factor of exponential lr_scheduler')

  79.     parser.add_argument('--eta_min', type=float, default=0., help='eta_min in cosine_annealing scheduler')

  80.     parser.add_argument('--T_max', type=int, default=80, help='T-max in cosine_annealing scheduler')

  81.     parser.add_argument('--max_epoch', type=int, default=80, help='max epoch num to train the model')

  82.     parser.add_argument('--warmup_epochs', default=0, type=float, help='warmup epoch')

  83.     parser.add_argument('--weight_decay', type=float, default=0.001, help='weight decay')

  84.     parser.add_argument('--momentum', type=float, default=0.98, help='momentum')

  85. 

  86.     # logging related

  87.     parser.add_argument('--log_interval', type=int, default=100, help='logging interval')

  88.     parser.add_argument('--ckpt_path', type=str, default='train_outputs/', help='checkpoint save location')

  89.     parser.add_argument('--ckpt_interval', type=int, default=100, help='save ckpt_interval')

  90. 

  91.     flags, _ = parser.parse_known_args()

  92.     flags.dataset_sink_mode = bool(flags.use_graph_mode)

  93.     flags.lr_epochs = list(map(int, flags.lr_epochs.split(',')))

  94. 

  95.     model_settings = prepare_model_settings(

  96.         len(prepare_words_list(flags.wanted_words.split(','))),

  97.         flags.sample_rate, flags.clip_duration_ms, flags.window_size_ms,

  98.         flags.window_stride_ms, flags.dct_coefficient_count)

  99.     model_settings['dropout1'] = flags.drop

  100.     return flags, model_settings

  101. 

  102. 

  103. def eval_config(parser):

  104.     '''config for eval.'''

  105.     parser.add_argument('--feat_dir', type=str, default='feat',

  106.                         help='Where to save the feature of audios')

  107.     parser.add_argument('--model_dir', type=str,

  108.                         default='outputs',

  109.                         help='which folder the models are saved in or specific path of one model')

  110.     parser.add_argument('--wanted_words', type=str, default='yes,no,up,down,left,right,on,off,stop,go',

  111.                         help='Words to use (others will be added to an unknown label)')

  112.     parser.add_argument('--sample_rate', type=int, default=16000, help='Expected sample rate of the wavs')

  113.     parser.add_argument('--clip_duration_ms', type=int, default=1000,

  114.                         help='Expected duration in milliseconds of the wavs')

  115.     parser.add_argument('--window_size_ms', type=float, default=40.0, help='How long each spectrogram timeslice is')

  116.     parser.add_argument('--window_stride_ms', type=float, default=20.0, help='How long each spectrogram timeslice is')

  117.     parser.add_argument('--dct_coefficient_count', type=int, default=20,

  118.                         help='How many bins to use for the MFCC fingerprint')

  119.     parser.add_argument('--model_size_info', type=int, nargs="+",

  120.                         default=[6, 276, 10, 4, 2, 1, 276, 3, 3, 2, 2, 276, 3, 3, 1, 1, 276, 3, 3, 1, 1, 276, 3, 3, 1,

  121.                                  1, 276, 3, 3, 1, 1],

  122.                         help='Model dimensions - different for various models')

  123. 

  124.     parser.add_argument('--per_batch_size', default=100, type=int, help='batch size for per gpu')

  125.     parser.add_argument('--drop', type=float, default=0.9, help='dropout')

  126. 

  127.     # logging related

  128.     parser.add_argument('--log_path', type=str, default='eval_outputs/', help='path to save eval log')

  129. 

  130.     flags, _ = parser.parse_known_args()

  131.     model_settings = prepare_model_settings(

  132.         len(prepare_words_list(flags.wanted_words.split(','))),

  133.         flags.sample_rate, flags.clip_duration_ms, flags.window_size_ms,

  134.         flags.window_stride_ms, flags.dct_coefficient_count)

  135.     model_settings['dropout1'] = flags.drop

  136.     return flags, model_settings

  137. 

  138. 

  139. def prepare_model_settings(label_count, sample_rate, clip_duration_ms,

  140.                            window_size_ms, window_stride_ms,

  141.                            dct_coefficient_count):

  142.     '''Prepare model setting.'''

  143.     desired_samples = int(sample_rate * clip_duration_ms / 1000)

  144.     window_size_samples = int(sample_rate * window_size_ms / 1000)

  145.     window_stride_samples = int(sample_rate * window_stride_ms / 1000)

  146.     length_minus_window = (desired_samples - window_size_samples)

  147.     if length_minus_window < 0:

  148.         spectrogram_length = 0

  149.     else:

  150.         spectrogram_length = 1 + int(length_minus_window / window_stride_samples)

  151.     fingerprint_size = dct_coefficient_count * spectrogram_length

  152.     return {

  153.         'desired_samples': desired_samples,

  154.         'window_size_samples': window_size_samples,

  155.         'window_stride_samples': window_stride_samples,

  156.         'spectrogram_length': spectrogram_length,

  157.         'dct_coefficient_count': dct_coefficient_count,

  158.         'fingerprint_size': fingerprint_size,

  159.         'label_count': label_count,

  160.         'sample_rate': sample_rate,

  161.     }