doka-noclip-gln-sgd.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
@@ -59,7 +59,7 @@ 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')
@@ -67,7 +67,7 @@ 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)')
@@ -107,7 +107,8 @@ 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")
@@ -151,8 +152,8 @@ 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(2)
+                                    shuffle=True, num_shards=rank_size, shard_id=rank_id)
+    tr_loader = tr_loader.batch(1)
     num_steps = tr_loader.get_dataset_size()
 
     # build model
@@ -171,7 +172,7 @@ 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='commit',
                               directory=save_ckpt,
                               config=config_ck)

test.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)')
@@ -110,8 +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")
@@ -162,7 +161,7 @@ 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()
     end_time = time.perf_counter()
@@ -184,7 +183,7 @@ 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='commit',
                               directory=save_ckpt,
                               config=config_ck)

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=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=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)')
@@ -110,7 +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")
@@ -161,7 +162,7 @@ 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=False, num_shards=rank_size, shard_id=rank_id)
+                                    shuffle=True, num_shards=rank_size, shard_id=rank_id)
     tr_loader = tr_loader.batch(4)
     num_steps = tr_loader.get_dataset_size()
     end_time = time.perf_counter()
@@ -183,7 +184,7 @@ 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='commit',
                               directory=save_ckpt,
                               config=config_ck)