Merge pull request sooftware#139 from hwiorn/fix-cpu-training-for-con…

…former

Fix cpu training for conformer

bin/main.py

@@ -77,6 +77,8 @@ def train(config: DictConfig) -> nn.DataParallel:
     torch.manual_seed(config.train.seed)
     torch.cuda.manual_seed_all(config.train.seed)
     device = check_envirionment(config.train.use_cuda)
+    if config.train.num_threads and int(config.train.num_threads) > 0:
+        torch.set_num_threads(config.train.num_threads)
 
     vocab = KsponSpeechVocabulary(
         f'../../../data/vocab/aihub_{config.train.output_unit}_vocabs.csv',

kospeech/model_builder.py

@@ -222,7 +222,7 @@ def build_conformer(
         half_step_residual=half_step_residual,
         device=device,
         decoder=decoder,
-    ))
+    )).to(device)
 
 
 def build_deepspeech2(

kospeech/models/conformer/encoder.py

@@ -91,6 +91,7 @@ def __init__(
                     d_model=encoder_dim,
                     num_heads=num_attention_heads,
                     dropout_p=attention_dropout_p,
+                    device=device,
                 ),
             ),
             ResidualConnectionModule(
@@ -99,13 +100,15 @@ def __init__(
                     kernel_size=conv_kernel_size,
                     expansion_factor=conv_expansion_factor,
                     dropout_p=conv_dropout_p,
+                    device=device,
                 ),
             ),
             ResidualConnectionModule(
                 module=FeedForwardModule(
                     encoder_dim=encoder_dim,
                     expansion_factor=feed_forward_expansion_factor,
                     dropout_p=feed_forward_dropout_p,
+                    device=device,
                 ),
                 module_factor=self.feed_forward_residual_factor,
             ),

kospeech/trainer/__init__.py

@@ -32,6 +32,7 @@ class TrainConfig:
 
     num_workers: int = 4
     use_cuda: bool = True
+    num_threads: int = 2
 
     init_lr_scale: float = 0.01
     final_lr_scale: float = 0.05