diff --git a/train_asteroid.py b/train_asteroid.py index c61a4af..e2fab24 100644 --- a/train_asteroid.py +++ b/train_asteroid.py @@ -42,7 +42,7 @@ parser.add_argument('--sample_rate', default=8000, type=int, help='Sample rate') parser.add_argument('--segment', default=4, type=float, # 取音频的长度，2s。#数据集语音长度要相同 help='Segment length (seconds)') -parser.add_argument('--batch_size', default=3, type=int, # 需要抛弃的音频长度 +parser.add_argument('--batch_size', default=2, type=int, # 需要抛弃的音频长度 help='Batch size') # Network architecture @@ -62,7 +62,7 @@ parser.add_argument('--norm', default='gln', type=str, help='gln = "Global Norm", cln = "Cumulative Norm", ln = "Layer Norm"') parser.add_argument('--dropout', default=0.0, type=float, help='dropout') -parser.add_argument('--num_layers', default=4, type=int, +parser.add_argument('--num_layers', default=6, type=int, help='Number of Dual-Path-Block') parser.add_argument('--K', default=250, type=int, help='The length of chunk') @@ -70,7 +70,7 @@ parser.add_argument('--num_spks', default=2, type=int, help='The number of speakers') # optimizer -parser.add_argument('--lr', default=0.001, type=float, +parser.add_argument('--lr', default=1e-3, type=float, help='Init learning rate') parser.add_argument('--l2', default=1e-5, type=float, help='weight decay (L2 penalty)') @@ -110,8 +110,7 @@ def preprocess(args): print("preprocess done") def main(args): - # context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target, save_graphs=True) - context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) + context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target, save_graphs=True) if args.run_distribute: print("distribute") @@ -162,8 +161,8 @@ def main(args): 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, num_shards=rank_size, shard_id=rank_id) - tr_loader = tr_loader.batch(4) + shuffle=True, num_shards=rank_size, shard_id=rank_id) + tr_loader = tr_loader.batch(2) num_steps = tr_loader.get_dataset_size() end_time = time.perf_counter() print("preparing data use: {}min".format((end_time - start_time) / 60)) @@ -187,7 +186,7 @@ def main(args): loss_cb = LossMonitor(1) cb = [time_cb, loss_cb] - config_ck = CheckpointConfig(save_checkpoint_steps=200, keep_checkpoint_max=5) + config_ck = CheckpointConfig(save_checkpoint_steps=100, keep_checkpoint_max=5) ckpt_cb = ModelCheckpoint(prefix='DPRNN', directory=save_ckpt, config=config_ck) diff --git a/train_clipnorm_new.py b/train_clipnorm_new.py index afba65b..534b4a1 100644 --- a/train_clipnorm_new.py +++ b/train_clipnorm_new.py @@ -42,7 +42,7 @@ parser.add_argument('--sample_rate', default=8000, type=int, help='Sample rate') parser.add_argument('--segment', default=4, type=float, # 取音频的长度，2s。#数据集语音长度要相同 help='Segment length (seconds)') -parser.add_argument('--batch_size', default=2, type=int, # 需要抛弃的音频长度 +parser.add_argument('--batch_size', default=3, type=int, # 需要抛弃的音频长度 help='Batch size') # Network architecture @@ -62,7 +62,7 @@ parser.add_argument('--norm', default='gln', type=str, help='gln = "Global Norm", cln = "Cumulative Norm", ln = "Layer Norm"') parser.add_argument('--dropout', default=0.0, type=float, help='dropout') -parser.add_argument('--num_layers', default=6, type=int, +parser.add_argument('--num_layers', default=4, type=int, help='Number of Dual-Path-Block') parser.add_argument('--K', default=250, type=int, help='The length of chunk') @@ -70,7 +70,7 @@ parser.add_argument('--num_spks', default=2, type=int, help='The number of speakers') # optimizer -parser.add_argument('--lr', default=1e-3, type=float, +parser.add_argument('--lr', default=0.001, type=float, help='Init learning rate') parser.add_argument('--l2', default=1e-5, type=float, help='weight decay (L2 penalty)') @@ -110,7 +110,8 @@ def preprocess(args): print("preprocess done") def main(args): - context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target, save_graphs=True) + # context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target, save_graphs=True) + context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) if args.run_distribute: print("distribute") @@ -159,7 +160,7 @@ def main(args): 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=True, num_shards=rank_size, shard_id=rank_id) + shuffle=False, num_shards=rank_size, shard_id=rank_id) tr_loader = tr_loader.batch(4) num_steps = tr_loader.get_dataset_size() @@ -184,7 +185,7 @@ def main(args): loss_cb = LossMonitor(1) cb = [time_cb, loss_cb] - config_ck = CheckpointConfig(save_checkpoint_steps=100, keep_checkpoint_max=5) + config_ck = CheckpointConfig(save_checkpoint_steps=200, keep_checkpoint_max=5) ckpt_cb = ModelCheckpoint(prefix='DPRNN', directory=save_ckpt, config=config_ck) diff --git a/train_ln_adam.py b/train_ln_adam.py index a7f4df0..330ead2 100644 --- a/train_ln_adam.py +++ b/train_ln_adam.py @@ -41,7 +41,7 @@ parser.add_argument('--sample_rate', default=8000, type=int, help='Sample rate') parser.add_argument('--segment', default=4, type=float, # 取音频的长度，2s。#数据集语音长度要相同 help='Segment length (seconds)') -parser.add_argument('--batch_size', default=3, type=int, # 需要抛弃的音频长度 +parser.add_argument('--batch_size', default=2, type=int, # 需要抛弃的音频长度 help='Batch size') # Network architecture @@ -61,7 +61,7 @@ parser.add_argument('--norm', default='gln', type=str, help='gln = "Global Norm", cln = "Cumulative Norm", ln = "Layer Norm"') parser.add_argument('--dropout', default=0.0, type=float, help='dropout') -parser.add_argument('--num_layers', default=4, type=int, +parser.add_argument('--num_layers', default=6, type=int, help='Number of Dual-Path-Block') parser.add_argument('--K', default=250, type=int, help='The length of chunk') @@ -69,7 +69,7 @@ parser.add_argument('--num_spks', default=2, type=int, help='The number of speakers') # optimizer -parser.add_argument('--lr', default=0.001, type=float, +parser.add_argument('--lr', default=1e-3, type=float, help='Init learning rate') parser.add_argument('--l2', default=1e-5, type=float, help='weight decay (L2 penalty)') @@ -109,7 +109,8 @@ def preprocess(args): print("preprocess done") def main(args): - context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target, save_graphs=True) + # context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target, save_graphs=True) + context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target) if args.run_distribute: print("distribute") @@ -154,7 +155,7 @@ def main(args): 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=True, num_shards=rank_size, shard_id=rank_id) + shuffle=False, num_shards=rank_size, shard_id=rank_id) tr_loader = tr_loader.batch(4) num_steps = tr_loader.get_dataset_size() # build model @@ -172,7 +173,7 @@ def main(args): loss_cb = LossMonitor(1) cb = [time_cb, loss_cb] - config_ck = CheckpointConfig(save_checkpoint_steps=200, keep_checkpoint_max=5) + config_ck = CheckpointConfig(save_checkpoint_steps=100, keep_checkpoint_max=5) ckpt_cb = ModelCheckpoint(prefix='DPRNN', directory=save_ckpt, config=config_ck)