From 7b815a37c32d1ed5f382e415f501cdefed19f0ad Mon Sep 17 00:00:00 2001 From: unicorn <15684175528@163.com> Date: Thu, 1 Dec 2022 21:50:10 +0800 Subject: [PATCH 1/5] =?UTF-8?q?=E6=9B=B4=E6=96=B0=20'train1.py'?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- train1.py | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/train1.py b/train1.py index f1a3262..0230531 100644 --- a/train1.py +++ b/train1.py @@ -40,7 +40,7 @@ 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 @@ -66,7 +66,7 @@ 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)') @@ -96,7 +96,7 @@ def preprocess(args): print("preprocess done") def main(args): - 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) ######################## 将数据集从obs拷贝到训练镜像中 （固定写法）######################## home = os.path.dirname(os.path.realpath(__file__)) @@ -125,8 +125,8 @@ def main(args): # build dataloader 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(batch_size=2) + tr_loader = ds.GeneratorDataset(tr_dataset, ["mixture", "lens", "sources"], shuffle=True) + tr_loader = tr_loader.batch(batch_size=1) num_steps = tr_loader.get_dataset_size() end_time = time.perf_counter() print("preparing data use: {}min".format((end_time - start_time) / 60)) @@ -147,7 +147,7 @@ def main(args): time_cb = TimeMonitor() cb = [time_cb, loss_cb] - config_ck = CheckpointConfig(save_checkpoint_steps=num_steps, keep_checkpoint_max=5) + config_ck = CheckpointConfig(save_checkpoint_steps=200, keep_checkpoint_max=5) ckpt_cb = ModelCheckpoint(prefix='DPRNN', directory=args.save_checkpoint_path, config=config_ck) @@ -155,7 +155,7 @@ def main(args): #开始训练 print("============== Starting Training ==============") - model.train(epoch=10, train_dataset=tr_loader, callbacks=cb, dataset_sink_mode=False) + model.train(epoch=1, train_dataset=tr_loader, callbacks=cb, dataset_sink_mode=False) ######################## 将输出的模型拷贝到obs（固定写法） ######################## try: -- 2.34.1 From b4728fdb2e763b4e1ce73da789311b43c6d22b6a Mon Sep 17 00:00:00 2001 From: unicorn <15684175528@163.com> Date: Thu, 1 Dec 2022 23:15:30 +0800 Subject: [PATCH 2/5] =?UTF-8?q?=E6=9B=B4=E6=96=B0=20'model=5Frnn1.py'?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- model_rnn1.py | 22 +++++++++++----------- 1 file changed, 11 insertions(+), 11 deletions(-) diff --git a/model_rnn1.py b/model_rnn1.py index 039c6d7..688f9a5 100644 --- a/model_rnn1.py +++ b/model_rnn1.py @@ -217,12 +217,12 @@ class Dual_Path_RNN(nn.Cell): def construct(self, x): # [B, N, L] - # if(self.selectNorm == 'ln'): - # x = self.expand_dims(x, 0).transpose((0, 2, 1, 3)) - # x = self.norm1(x) - # x = x.transpose((0, 2, 1, 3)).squeeze(axis=0) - # else: - x = self.norm(x) + if(self.selectNorm == 'ln'): + x = self.expand_dims(x, 0).transpose((0, 2, 1, 3)) + x = self.norm1(x) + x = x.transpose((0, 2, 1, 3)).squeeze(axis=0) + else: + x = self.norm(x) # [B, N, L] x = self.conv1d(x) # [B, N, K, S] @@ -325,9 +325,9 @@ class Dual_RNN_model(nn.Cell): self.print = ops.Print() self.stack = ops.Stack() - for p in self.get_parameters(): - if p.ndim > 1: - mindspore.common.initializer.HeNormal(p) + # for p in self.get_parameters(): + # if p.ndim > 1: + # mindspore.common.initializer.HeNormal(p) def construct(self, x): # [B, N, L] @@ -344,8 +344,8 @@ class Dual_RNN_model(nn.Cell): if __name__ == "__main__": context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=0) - rnn = Dual_RNN_model(256, 64, 128, bidirectional=True, norm='ln', num_layers=6, dropout=0.0) + rnn = Dual_RNN_model(256, 64, 128, bidirectional=True, norm='gln', num_layers=4, dropout=0.0) ones = ops.Ones() - x = ones((2, 32000), mindspore.float32) + x = ones((1, 32000), mindspore.float32) out = rnn(x) print(rnn) -- 2.34.1 From 359a025a2a69aa8200130928719b5abb0de43a74 Mon Sep 17 00:00:00 2001 From: unicorn <15684175528@163.com> Date: Fri, 2 Dec 2022 23:20:09 +0800 Subject: [PATCH 3/5] =?UTF-8?q?=E6=9B=B4=E6=96=B0=20'train=5Fasteroid.py'?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- train_asteroid.py | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-) 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) -- 2.34.1 From 087ff36d1f57c2d3ce995779e18e8cf7e539269e Mon Sep 17 00:00:00 2001 From: unicorn <15684175528@163.com> Date: Sat, 3 Dec 2022 11:31:40 +0800 Subject: [PATCH 4/5] =?UTF-8?q?=E6=9B=B4=E6=96=B0=20'train=5Fclipnorm=5Fne?= =?UTF-8?q?w.py'?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- train_clipnorm_new.py | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-) 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) -- 2.34.1 From 40ad25f5ecb250db84992c92811d363c5ac59437 Mon Sep 17 00:00:00 2001 From: unicorn <15684175528@163.com> Date: Sat, 3 Dec 2022 21:54:58 +0800 Subject: [PATCH 5/5] =?UTF-8?q?=E6=9B=B4=E6=96=B0=20'train=5Fln=5Fadam.py'?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- train_ln_adam.py | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-) 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) -- 2.34.1