|
- import argparse
-
- from model_final import DPTNet_base
- from mindspore import Model
- from data_loader import DatasetGenerator
-
- import mindspore.dataset as ds
- from mindspore import nn, context
- from mindspore.train.callback import LossMonitor, TimeMonitor, ModelCheckpoint, CheckpointConfig
- from network_define import WithLossCell
- from Loss_final import Loss
-
-
- parser = argparse.ArgumentParser(
- "Dual-path transformer"
- "with Permutation Invariant Training")
- # General config
- # Task related
-
- # parser.add_argument('--train_dir', type=str, default='/home/ma-user/work/DPTNet_mindspore/out_dir/tr',
- # help='directory including mix.json, s1.json and s2.json')
- # parser.add_argument('--valid_dir', type=str, default='/home/ma-user/work/DPTNet_mindspore/out_dir/cv',
- # help='directory including mix.json, s1.json and s2.json')
- parser.add_argument('--train_dir', type=str, default='/home/ma-user/work/DPTNet_mindspore/out_dir/tr',
- help='directory including mix.json, s1.json and s2.json')
- parser.add_argument('--valid_dir', type=str, default='/home/ma-user/work/DPTNet_mindspore/out_dir/cv',
- help='directory including mix.json, s1.json and s2.json')
- parser.add_argument('--sample_rate', default=8000, type=int,
- help='Sample rate')
- parser.add_argument('--segment', default=4, type=float,
- help='Segment length (seconds)')
- parser.add_argument('--cv_maxlen', default=8, type=float,
- help='max audio length (seconds) in cv, to avoid OOM issue.')
- # Network architecture
- parser.add_argument('--N', default=64, type=int,
- help='Number of filters in autoencoder')
- parser.add_argument('--C', default=2, type=int,
- help='Maximum number of speakers')
- parser.add_argument('--L', default=4, type=int,
- help='Length of window in autoencoder') # L=2 in paper
- parser.add_argument('--H', default=4, type=int,
- help='Number of head in Multi-head attention')
- parser.add_argument('--K', default=250, type=int,
- help='segment size')
- parser.add_argument('--B', default=6, type=int,
- help='Number of repeats')
-
- parser.add_argument('--enc_dim', default=256, type=int,
- help='...')
- parser.add_argument('--feature_dim', default=64, type=int,
- help='Number of filters in autoencoder')
- parser.add_argument('--hidden_dim', default=128, type=int,
- help='...')
- parser.add_argument('--layer', default=6, type=int,
- help='Number of repeats')
- parser.add_argument('--segment_size', default=250, type=int,
- help='segment size')
- parser.add_argument('--nspk', default=2, type=int,
- help='Maximum number of speakers')
- parser.add_argument('--win_len', default=1, type=int,
- help='...')
-
- # Training config
- parser.add_argument('--use_cuda', type=int, default=1,
- help='Whether use GPU')
- parser.add_argument('--epochs', default=100, type=int,
- help='Number of maximum epochs')
- parser.add_argument('--half_lr', dest='half_lr', default=0, type=int,
- help='Halving learning rate when get small improvement')
- parser.add_argument('--early_stop', dest='early_stop', default=0, type=int,
- help='Early stop training when no improvement for 10 epochs')
- parser.add_argument('--max_norm', default=5, type=float,
- help='Gradient norm threshold to clip')
- # minibatch
- parser.add_argument('--shuffle', default=0, type=int,
- help='reshuffle the data at every epoch')
- parser.add_argument('--batch_size', default=2, type=int, #default =3
- help='Batch size')
- parser.add_argument('--num_workers', default=4, type=int, #default = 8
- help='Number of workers to generate minibatch')
- # optimizer
- parser.add_argument('--optimizer', default='adam', type=str,
- choices=['sgd', 'adam'],
- help='Optimizer (support sgd and adam now)')
- parser.add_argument('--lr', default=4e-4, type=float,
- help='Init learning rate')
- parser.add_argument('--momentum', default=0.0, type=float,
- help='Momentum for optimizer')
- parser.add_argument('--l2', default=0.0, type=float,
- help='weight decay (L2 penalty)')
- # save and load model
- parser.add_argument('--save_folder', default='exp/temp',
- help='Location to save epoch models')
- parser.add_argument('--checkpoint', dest='checkpoint', default=0, type=int,
- help='Enables checkpoint saving of model')
- parser.add_argument('--continue_from', default='',
- help='Continue from checkpoint model')
- parser.add_argument('--model_path', default='final.pth.tar',
- help='Location to save best validation model')
- # logging
- parser.add_argument('--print_freq', default=1000, type=int,
- help='Frequency of printing training infomation')
- parser.add_argument('--visdom', dest='visdom', type=int, default=0,
- help='Turn on visdom graphing')
- parser.add_argument('--visdom_epoch', dest='visdom_epoch', type=int, default=0,
- help='Turn on visdom graphing each epoch')
- parser.add_argument('--visdom_id', default='TasNet training',
- help='Identifier for visdom run')
-
-
- def main(args):
- # Construct Solver
- # data
- tr_dataset = DatasetGenerator(args.train_dir, args.batch_size,
- sample_rate=args.sample_rate, segment=args.segment)
- tr_loader = ds.GeneratorDataset(tr_dataset, ["mixture", "lens", "sources"], shuffle=False)
- tr_loader = tr_loader.batch(2)
- # model
- net = DPTNet_base(enc_dim=256, feature_dim=64, hidden_dim=128, layer=1, segment_size=250, nspk=2, win_len=2)
-
- net.set_train()
- optimizier = nn.Adam(net.get_parameters(), learning_rate=args.lr, weight_decay=args.l2)
- my_loss = Loss()
- net_with_loss = WithLossCell(net, my_loss)
- model = Model(net_with_loss, optimizer=optimizier)
-
- time_cb = TimeMonitor()
- loss_cb = LossMonitor(10)
- cb = [time_cb, loss_cb]
- config_ck = CheckpointConfig(save_checkpoint_steps=5,
- keep_checkpoint_max=5)
- ckpt_cb = ModelCheckpoint(prefix="DPTNet", directory=args.save_folder, config=config_ck)
-
- cb += [ckpt_cb]
-
- model.train(epoch=200, train_dataset=tr_loader, callbacks=cb, dataset_sink_mode=False)
-
- if __name__ == '__main__':
- context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=0)
- # context.set_context(mode=context.PYNATIVE_MODE, device_target="Ascend", device_id=7)
- # context.set_context(mode=context.GRAPH_MODE, device_target="GPU", device_id=1)
- # context.set_context(mode=context.PYNATIVE_MODE, device_target="GPU", device_id=1)
- args = parser.parse_args()
- print(args)
- main(args)
|