update

examples/nx_denoise/yaml/config.yaml

@@ -1,11 +1,20 @@
-model_name: "nx_clean_unet"
+model_name: "nx_denoise"
 
 sample_rate: 8000
 segment_size: 16000
 n_fft: 512
 win_size: 200
 hop_size: 80
+# 因为 hop_size 取 80，则相当于 stft 的时间步是 10ms 一步，所以降采样也考虑到差不多的分辨率。
 
+# 2**down_sampling_num_layers，
+# 例如 2**6=64 就意味着 64 个值在降采样之后是一个时间步，
+# 则一步是 64/sample_rate = 0.008秒。
+# 那么 tsfm_chunk_size=2 则为16ms，tsfm_chunk_size=4 则为32ms
+# 假设每次向左看1秒，向右看30ms，则：
+# tsfm_chunk_size=1，tsfm_num_left_chunks=128，tsfm_num_right_chunks=4
+# tsfm_chunk_size=2，tsfm_num_left_chunks=64，tsfm_num_right_chunks=2
+# tsfm_chunk_size=4，tsfm_num_left_chunks=32，tsfm_num_right_chunks=1
 down_sampling_num_layers: 6
 down_sampling_in_channels: 1
 down_sampling_hidden_channels: 64
@@ -13,7 +22,7 @@ down_sampling_kernel_size: 4
 down_sampling_stride: 2
 
 causal_in_channels: 1
-causal_out_channels: 1
+causal_out_channels: 64
 causal_kernel_size: 3
 causal_bias: false
 causal_separable: true
@@ -34,7 +43,7 @@ discriminator_in_channel: 2
 
 compress_factor: 0.3
 
-batch_size: 64
+batch_size: 4
 learning_rate: 0.0005
 adam_b1: 0.8
 adam_b2: 0.99

toolbox/torchaudio/models/nx_denoise/yaml/config.yaml

@@ -1,4 +1,4 @@
-model_name: "nx_clean_unet"
+model_name: "nx_denoise"
 
 sample_rate: 8000
 segment_size: 16000