update

examples/nx_mpnet/yaml/config.yaml

@@ -19,8 +19,8 @@ tsfm_hidden_size: 64
 tsfm_attention_heads: 4
 tsfm_num_blocks: 4
 tsfm_dropout_rate: 0.0
-tsfm_max_time_relative_position: 1024
-tsfm_max_freq_relative_position: 128
+tsfm_max_time_relative_position: 2048
+tsfm_max_freq_relative_position: 256
 tsfm_chunk_size: 1
 tsfm_num_left_chunks: 64
 tsfm_num_right_chunks: 2

toolbox/torchaudio/models/nx_mpnet/configuration_nx_mpnet.py

@@ -27,8 +27,8 @@ class NXMPNetConfig(PretrainedConfig):
                  tsfm_attention_heads: int = 4,
                  tsfm_num_blocks: int = 4,
                  tsfm_dropout_rate: float = 0.0,
-                 tsfm_max_time_relative_position: int = 1024,
-                 tsfm_max_freq_relative_position: int = 128,
+                 tsfm_max_time_relative_position: int = 2048,
+                 tsfm_max_freq_relative_position: int = 256,
                  tsfm_chunk_size: int = 1,
                  tsfm_num_left_chunks: int = 64,
                  tsfm_num_right_chunks: int = 2,

toolbox/torchaudio/models/nx_mpnet/transformers/transformers.py

@@ -245,8 +245,8 @@ class TSTransformerBlock(nn.Module):
                  input_dim: int,
                  dropout_rate: float = 0.1,
                  n_heads: int = 4,
-                 max_time_relative_position: int = 1024,
-                 max_freq_relative_position: int = 128,
+                 max_time_relative_position: int = 2048,
+                 max_freq_relative_position: int = 256,
                  ):
         super(TSTransformerBlock, self).__init__()
         self.time_transformer = TransformerBlock(input_dim, dropout_rate, n_heads, max_time_relative_position)
@@ -285,8 +285,8 @@ class TSTransformerEncoder(nn.Module):
                  attention_heads: int = 4,
                  num_blocks: int = 6,
                  dropout_rate: float = 0.1,
-                 max_time_relative_position: int = 1024,
-                 max_freq_relative_position: int = 128,
+                 max_time_relative_position: int = 2048,
+                 max_freq_relative_position: int = 256,
                  chunk_size: int = 1,
                  num_left_chunks: int = 128,
                  num_right_chunks: int = 2,

toolbox/torchaudio/models/nx_mpnet/yaml/config.yaml

@@ -19,8 +19,8 @@ tsfm_hidden_size: 64
 tsfm_attention_heads: 4
 tsfm_num_blocks: 4
 tsfm_dropout_rate: 0.0
-tsfm_max_time_relative_position: 1024
-tsfm_max_freq_relative_position: 128
+tsfm_max_time_relative_position: 2048
+tsfm_max_freq_relative_position: 256
 tsfm_chunk_size: 1
 tsfm_num_left_chunks: 64
 tsfm_num_right_chunks: 2