update

.gitignore

@@ -1,4 +1,5 @@
 
+.gradio/
 .git/
 .idea/
 

README.md

@@ -12,7 +12,7 @@ Check out the configuration reference at https://huggingface.co/docs/hub/spaces-
 ## NX Denoise
 
 
-### speech datasets
+### datasets
 
 ```text
 
@@ -22,5 +22,8 @@ https://openslr.trmal.net/resources/33/
 AISHELL-3 (19G)
 http://www.openslr.org/93/
 
+DNS3
+https://github.com/microsoft/DNS-Challenge/blob/master/download-dns-challenge-3.sh
+
 ```
 

examples/frcrn/run.sh

@@ -4,7 +4,8 @@
 
 
 sh run.sh --stage 2 --stop_stage 2 --system_version centos --file_folder_name file_dir --final_model_name frcrn \
---noise_dir "/data/tianxing/HuggingDatasets/nx_noise/data/noise/dns3-noise" \
+--config_file "yaml/config-20-512.yaml" \
+--noise_dir "/data/tianxing/HuggingDatasets/nx_noise/data/noise" \
 --speech_dir "/data/tianxing/HuggingDatasets/nx_noise/data/speech/dns3-speech"
 
 

examples/frcrn/yaml/{config.yaml → config-10-512.yaml}

@@ -1,24 +1,10 @@
-
 model_name: "frcrn"
 
-num_gpus: -1
-
-lr: 0.001
-lr_scheduler: "CosineAnnealingLR"
-lr_scheduler_kwargs:
-  T_max: 250000
-  eta_min: 0.0001
-
-max_epochs: 100
-weight_decay: 1.0e-05
-clip_grad_norm: 10.0
-seed: 1234
-
 sample_rate: 8000
 segment_size: 32000
-nfft: 128
-win_size: 128
-hop_size: 64
+nfft: 512
+win_size: 512
+hop_size: 128
 win_type: hann
 
 use_complex_networks: true
@@ -31,3 +17,15 @@ max_snr_db: 20
 num_workers: 8
 batch_size: 32
 eval_steps: 10000
+
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+weight_decay: 1.0e-05
+clip_grad_norm: 10.0
+seed: 1234
+num_gpus: -1

examples/frcrn/yaml/config-14-512.yaml

@@ -0,0 +1,31 @@
+model_name: "frcrn"
+
+sample_rate: 8000
+segment_size: 32000
+nfft: 512
+win_size: 512
+hop_size: 128
+win_type: hann
+
+use_complex_networks: true
+model_depth: 14
+model_complexity: -1
+
+min_snr_db: -10
+max_snr_db: 20
+
+num_workers: 8
+batch_size: 32
+eval_steps: 10000
+
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+weight_decay: 1.0e-05
+clip_grad_norm: 10.0
+seed: 1234
+num_gpus: -1

examples/frcrn/yaml/config-20-512.yaml

@@ -0,0 +1,31 @@
+model_name: "frcrn"
+
+sample_rate: 8000
+segment_size: 32000
+nfft: 512
+win_size: 512
+hop_size: 128
+win_type: hann
+
+use_complex_networks: true
+model_depth: 20
+model_complexity: 45
+
+min_snr_db: -10
+max_snr_db: 20
+
+num_workers: 8
+batch_size: 32
+eval_steps: 10000
+
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+weight_decay: 1.0e-05
+clip_grad_norm: 10.0
+seed: 1234
+num_gpus: -1

main.py

@@ -1,14 +1,16 @@
 #!/usr/bin/python3
 # -*- coding: utf-8 -*-
 """
-docker build -t denoise:v20250609_1536 .
+docker build -t denoise:v20250609_1919 .
 docker stop denoise_7865 && docker rm denoise_7865
 docker run -itd \
 --name denoise_7865 \
 --restart=always \
 --network host \
--e port=7865 \
-denoise:v20250609_1536
+-e server_port=7865 \
+-e hf_token=hf_coRVvzwAzCwGHKRK***********EX \
+denoise:v20250609_1919 /bin/bash
+
 """
 import argparse
 import json
@@ -273,7 +275,9 @@ def main():
                 )
 
     # http://127.0.0.1:7865/
+    # http://10.75.27.247:7865/
     blocks.queue().launch(
+        # share=True,
         share=False if platform.system() == "Windows" else False,
         server_name="127.0.0.1" if platform.system() == "Windows" else "0.0.0.0",
         server_port=args.server_port

toolbox/torchaudio/models/frcrn/configuration_frcrn.py

@@ -10,22 +10,11 @@ from toolbox.torchaudio.configuration_utils import PretrainedConfig
 
 class FRCRNConfig(PretrainedConfig):
     def __init__(self,
-                 num_gpus: int = -1,
-
-                 lr: float = 0.001,
-                 lr_scheduler: str = "CosineAnnealingLR",
-                 lr_scheduler_kwargs: dict = None,
-
-                 max_epochs: int = 100,
-                 weight_decay: float = 0.00001,
-                 clip_grad_norm: float = 10.,
-                 seed: int = 1234,
-
                  sample_rate: int = 8000,
                  segment_size: int = 32000,
                  nfft: int = 512,
                  win_size: int = 512,
-                 hop_size: int = 256,
+                 hop_size: int = 128,
                  win_type: str = "hann",
 
                  use_complex_networks: bool = True,
@@ -39,20 +28,19 @@ class FRCRNConfig(PretrainedConfig):
                  batch_size: int = 4,
                  eval_steps: int = 25000,
 
+                 lr: float = 0.001,
+                 lr_scheduler: str = "CosineAnnealingLR",
+                 lr_scheduler_kwargs: dict = None,
+
+                 max_epochs: int = 100,
+                 weight_decay: float = 0.00001,
+                 clip_grad_norm: float = 10.,
+                 seed: int = 1234,
+                 num_gpus: int = -1,
+
                  **kwargs
                  ):
         super(FRCRNConfig, self).__init__(**kwargs)
-        self.num_gpus = num_gpus
-
-        self.lr = lr
-        self.lr_scheduler = lr_scheduler
-        self.lr_scheduler_kwargs = lr_scheduler_kwargs or dict()
-
-        self.max_epochs = max_epochs
-        self.weight_decay = weight_decay
-        self.clip_grad_norm = clip_grad_norm
-        self.seed = seed
-
         self.sample_rate = sample_rate
         self.segment_size = segment_size
         self.nfft = nfft
@@ -71,6 +59,16 @@ class FRCRNConfig(PretrainedConfig):
         self.batch_size = batch_size
         self.eval_steps = eval_steps
 
+        self.lr = lr
+        self.lr_scheduler = lr_scheduler
+        self.lr_scheduler_kwargs = lr_scheduler_kwargs or dict()
+
+        self.max_epochs = max_epochs
+        self.weight_decay = weight_decay
+        self.clip_grad_norm = clip_grad_norm
+        self.seed = seed
+        self.num_gpus = num_gpus
+
 
 def main():
     config = FRCRNConfig()

toolbox/torchaudio/models/frcrn/yaml/config-10-512.yaml

@@ -0,0 +1,31 @@
+model_name: "frcrn"
+
+sample_rate: 8000
+segment_size: 32000
+nfft: 512
+win_size: 512
+hop_size: 128
+win_type: hann
+
+use_complex_networks: true
+model_depth: 10
+model_complexity: -1
+
+min_snr_db: -10
+max_snr_db: 20
+
+num_workers: 8
+batch_size: 32
+eval_steps: 10000
+
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+weight_decay: 1.0e-05
+clip_grad_norm: 10.0
+seed: 1234
+num_gpus: -1

toolbox/torchaudio/models/frcrn/yaml/config-14-512.yaml

@@ -0,0 +1,31 @@
+model_name: "frcrn"
+
+sample_rate: 8000
+segment_size: 32000
+nfft: 512
+win_size: 512
+hop_size: 128
+win_type: hann
+
+use_complex_networks: true
+model_depth: 14
+model_complexity: -1
+
+min_snr_db: -10
+max_snr_db: 20
+
+num_workers: 8
+batch_size: 32
+eval_steps: 10000
+
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+weight_decay: 1.0e-05
+clip_grad_norm: 10.0
+seed: 1234
+num_gpus: -1

toolbox/torchaudio/models/frcrn/yaml/config-20-512.yaml

@@ -0,0 +1,31 @@
+model_name: "frcrn"
+
+sample_rate: 8000
+segment_size: 32000
+nfft: 512
+win_size: 512
+hop_size: 128
+win_type: hann
+
+use_complex_networks: true
+model_depth: 20
+model_complexity: 45
+
+min_snr_db: -10
+max_snr_db: 20
+
+num_workers: 8
+batch_size: 32
+eval_steps: 10000
+
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+weight_decay: 1.0e-05
+clip_grad_norm: 10.0
+seed: 1234
+num_gpus: -1

toolbox/torchaudio/models/tcnn/modeling_tcnn.py

@@ -233,12 +233,13 @@ class TCNN(nn.Module):
         if remainder > 0:
             n_samples_pad = self.hop_size - remainder
             signal = F.pad(signal, pad=(0, n_samples_pad), mode="constant", value=0)
-        return signal, n_samples
+        return signal
 
     def forward(self,
                 noisy: torch.Tensor,
                 ):
-        noisy, num_samples = self.signal_prepare(noisy)
+        num_samples = noisy.shape[-1]
+        noisy = self.signal_prepare(noisy)
         batch_size, _, num_samples_pad = noisy.shape
 
         # n_frame = (num_samples_pad - self.win_size) / self.hop_size + 1
@@ -268,6 +269,8 @@ class TCNN(nn.Module):
 
         denoise = denoise[:, :num_samples]
         # denoise shape: [b, num_samples]
+        denoise = torch.unsqueeze(denoise, dim=1)
+        # denoise shape: [b, 1, num_samples]
         return denoise
 
     def forward_chunk(self, inputs: torch.Tensor):
@@ -332,6 +335,7 @@ class TCNN(nn.Module):
 
 def main():
     model = TCNN()
+    model.eval()
 
     x = torch.randn(64, 1, 5, 320)
     # x = torch.randn(64, 1, 5, 160)