Merge pull request #27 from shahules786/dev

Rename dataset

README.md

@@ -30,13 +30,15 @@ model("noisy_audio.wav")
 
 | Model     | Dataset      | STOI    | PESQ  | URL                           |
 | :---:     |  :---:       | :---:   | :---: | :---:                         |
-| WaveUnet  | Vctk-28spk   | 0.836   | 2.78  |  shahules786/mayavoz-waveunet-valentini-28spk      |
+| WaveUnet  | Valentini-28spk   | 0.836   | 2.78  |  shahules786/mayavoz-waveunet-valentini-28spk      |
-| Demucs    | Vctk-28spk   | 0.961   | 2.56  |  shahules786/mayavoz-demucs-valentini-28spk       |
+| Demucs    | Valentini-28spk   | 0.961   | 2.56  |  shahules786/mayavoz-demucs-valentini-28spk       |
-| DCCRN     | Vctk-28spk   | 0.724   | 2.55  |  shahules786/mayavoz-dccrn-valentini-28spk         |
+| DCCRN     | Valentini-28spk   | 0.724   | 2.55  |  shahules786/mayavoz-dccrn-valentini-28spk         |
-| Demucs     | DNS2020 (20hrs)  | 0.56 | 1.26  |  shahules786/mayavoz-demucs-dns2020-20hr        |
+| Demucs     | MS-SNSD-20hrs  | 0.56 | 1.26  | shahules786/mayavoz-demucs-ms-snsd-20       |
 
 Test scores are based on respective test set associated with train dataset.
 
+**See [tutorials](/notebooks/) to train your custom model**
+
 ## Installation
 Only Python 3.8+ is officially supported (though it might work with Python 3.7)
 

mayavoz/cli/train_config/dataset/MS-SNSD.yaml

@@ -1,5 +1,5 @@
 _target_: mayavoz.data.dataset.MayaDataset
-name : dns-2020
+name : MS-SDSD
 root_dir : /Users/shahules/Myprojects/MS-SNSD
 duration : 2.0
 sampling_rate: 16000

mayavoz/cli/train_config/dataset/Valentini.yaml

@@ -1,5 +1,5 @@
 _target_: mayavoz.data.dataset.MayaDataset
-name : vctk
+name : Valentini
 root_dir : /scratch/c.sistc3/DS_10283_2791
 duration : 4.5
 stride : 2

mayavoz/cli/train_config/dataset/Vctk.yaml

@@ -1,13 +0,0 @@
-_target_: mayavoz.data.dataset.MayaDataset
-name : vctk
-root_dir : /scratch/c.sistc3/DS_10283_2791
-duration : 4.5
-stride : 2
-sampling_rate: 16000
-batch_size: 32
-valid_minutes : 15
-files:
-  train_clean : clean_trainset_28spk_wav
-  test_clean : clean_testset_wav
-  train_noisy : noisy_trainset_28spk_wav
-  test_noisy : noisy_testset_wav

mayavoz/data/fileprocessor.py

@@ -95,7 +95,7 @@ class Fileprocessor:
         if matching_function is None:
             if name.lower() in ("vctk", "valentini"):
                 return cls(clean_dir, noisy_dir, ProcessorFunctions.one_to_one)
-            elif name.lower() == "dns-2020":
+            elif name.lower() == "ms-snsd":
                 return cls(clean_dir, noisy_dir, ProcessorFunctions.one_to_many)
             else:
                 raise ValueError(

notebooks/Custom_model_training.ipynb

@@ -82,7 +82,7 @@
     "- `min_valid_minutes`: minimum validation in minutes. Validation is automatically selected from training set. (exclusive users).\n",
     "- `matching_function`: there are two types of mapping functions.\n",
     "    - `one_to_one` : In this one clean file will only have one corresponding noisy file. For example Valentini datasets\n",
-    "    - `one_to_many` : In this one clean file will only have one corresponding noisy file. For example DNS dataset.\n"
+    "    - `one_to_many` : In this one clean file will only have one corresponding noisy file. For example MS-SNSD dataset.\n"
    ]
   },
   {

notebooks/Getting_started.ipynb

@@ -302,7 +302,7 @@
    "source": [
     "Now there are two types of `matching_function`\n",
     "- `one_to_one` : In this one clean file will only have one corresponding noisy file. For example Valentini datasets\n",
-    "- `one_to_many` : In this one clean file will only have one corresponding noisy file. For example DNS dataset."
+    "- `one_to_many` : In this one clean file will only have one corresponding noisy file. For example MS-SNSD dataset."
    ]
   },
   {
@@ -479,7 +479,7 @@
    "metadata": {},
    "source": [
     "### TL;DR\n",
-    "Calling the following command would train mayavoz Demucs model on DNS-2020 dataset.\n",
+    "Calling the following command would train mayavoz Demucs model on MS-SNSD dataset.\n",
     "\n",
     "```bash\n",
     "mayavoz-train \\\n",
@@ -540,7 +540,7 @@
     "mayavoz-train --cfg job \\\n",
     "    model=Demucs \\\n",
     "    Demucs.sampling_rate=16000 \\\n",
-    "    dataset=DNS-2020\n",
+    "    dataset=MS-SNSD\n",
     "\n",
     "```\n",
     "\n",
@@ -562,7 +562,7 @@
     "```bash\n",
     "mayavoz-train \\\n",
     "    model=Demucs model.sampling_rate=16000 \\\n",
-    "    dataset=DNS-2020\n",
+    "    dataset=MS-SNSD\n",
     "\n",
     "```"
    ]
@@ -570,9 +570,9 @@
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "enhancer",
+   "display_name": "Python 3.8.13 ('enhancer')",
    "language": "python",
-   "name": "enhancer"
+   "name": "python3"
   },
   "language_info": {
    "codemirror_mode": {

recipes/DNS/DNS-2020/README.md

@@ -1,15 +0,0 @@
-### DNS Challenge's dataset
-
-The Deep Noise Suppression (DNS) Challenge is a single-channel speech enhancement
-challenge organized by Microsoft, with a focus on real-time applications.
-More info can be found on the [official page](https://dns-challenge.azurewebsites.net/).
-
-**References**
-The challenge paper, [here](https://arxiv.org/abs/2001.08662).
-```BibTex
-@misc{DNSChallenge2020,
-title={The INTERSPEECH 2020 Deep Noise Suppression Challenge: Datasets, Subjective Speech Quality and Testing Framework},
-author={Chandan K. A. Reddy and Ebrahim Beyrami and Harishchandra Dubey and Vishak Gopal and Roger Cheng and Ross Cutler and Sergiy Matusevych and Robert Aichner and Ashkan Aazami and Sebastian Braun and Puneet Rana and Sriram Srinivasan and Johannes Gehrke}, year={2020},
-doi=https://doi.org/10.48550/arXiv.2001.08662,
-}
-```

recipes/DNS/DNS-2020/cli/train_config/dataset/Vctk.yaml

@@ -1,13 +0,0 @@
-_target_: mayavoz.data.dataset.MayaDataset
-name : vctk
-root_dir : /scratch/c.sistc3/DS_10283_2791
-duration : 4.5
-stride : 2
-sampling_rate: 16000
-batch_size: 32
-valid_minutes : 15
-files:
-  train_clean : clean_trainset_28spk_wav
-  test_clean : clean_testset_wav
-  train_noisy : noisy_trainset_28spk_wav
-  test_noisy : noisy_testset_wav

recipes/DNS/DNS-2020/cli/train_config/hyperparameters/default.yaml

@@ -1,7 +0,0 @@
-loss : mae
-metric : [stoi,pesq,si-sdr]
-lr : 0.0003
-ReduceLr_patience : 5
-ReduceLr_factor : 0.2
-min_lr : 0.000001
-EarlyStopping_factor : 10

recipes/DNS/DNS-2020/cli/train_config/model/WaveUnet.yaml

@@ -1,5 +0,0 @@
-_target_: mayavoz.models.waveunet.WaveUnet
-num_channels : 1
-depth : 9
-initial_output_channels: 24
-sampling_rate : 16000

recipes/MS-SNSD/MS-SNSD-20/DCCRN/train.py

@@ -19,9 +19,9 @@ JOB_ID = os.environ.get("SLURM_JOBID", "0")
 
 
 @hydra.main(config_path="train_config", config_name="config")
-def main(config: DictConfig):
+def train(config: DictConfig):
 
-    OmegaConf.save(config, "config_log.yaml")
+    OmegaConf.save(config, "config.yaml")
 
     callbacks = []
     logger = MLFlowLogger(
@@ -96,7 +96,7 @@ def main(config: DictConfig):
     trainer.test(model)
 
     logger.experiment.log_artifact(
-        logger.run_id, f"{trainer.default_root_dir}/config_log.yaml"
+        logger.run_id, f"{trainer.default_root_dir}/config.yaml"
     )
 
     saved_location = os.path.join(
@@ -117,4 +117,4 @@ def main(config: DictConfig):
 
 
 if __name__ == "__main__":
-    main()
+    train()

recipes/MS-SNSD/MS-SNSD-20/DCCRN/train_config/config.yaml

@@ -1,6 +1,6 @@
 defaults:
   - model : Demucs
-  - dataset : Vctk
+  - dataset : MS-SNSD
   - optimizer : Adam
   - hyperparameters : default
   - trainer : default

recipes/MS-SNSD/MS-SNSD-20/DCCRN/train_config/dataset/MS-SNSD.yaml

@@ -1,10 +1,11 @@
 _target_: mayavoz.data.dataset.MayaDataset
+name : MS-SDSD
 root_dir : /Users/shahules/Myprojects/MS-SNSD
-name : dns-2020
+duration : 1.5
-duration : 2.0
+stride : 1
 sampling_rate: 16000
 batch_size: 32
-valid_size: 0.05
+min_valid_minutes: 25
 files:
   train_clean : CleanSpeech_training
   test_clean : CleanSpeech_training

recipes/MS-SNSD/MS-SNSD-20/DCCRN/train_config/hyperparameters/default.yaml

@@ -0,0 +1,7 @@
+loss : si-snr
+metric : [stoi,pesq]
+lr : 0.001
+ReduceLr_patience : 10
+ReduceLr_factor : 0.5
+min_lr : 0.000001
+EarlyStopping_factor : 10

recipes/MS-SNSD/MS-SNSD-20/Demucs/train.py

@@ -19,9 +19,9 @@ JOB_ID = os.environ.get("SLURM_JOBID", "0")
 
 
 @hydra.main(config_path="train_config", config_name="config")
-def main(config: DictConfig):
+def train(config: DictConfig):
 
-    OmegaConf.save(config, "config_log.yaml")
+    OmegaConf.save(config, "config.yaml")
 
     callbacks = []
     logger = MLFlowLogger(
@@ -96,7 +96,7 @@ def main(config: DictConfig):
     trainer.test(model)
 
     logger.experiment.log_artifact(
-        logger.run_id, f"{trainer.default_root_dir}/config_log.yaml"
+        logger.run_id, f"{trainer.default_root_dir}/config.yaml"
     )
 
     saved_location = os.path.join(
@@ -117,4 +117,4 @@ def main(config: DictConfig):
 
 
 if __name__ == "__main__":
-    main()
+    train()

recipes/MS-SNSD/MS-SNSD-20/Demucs/train_config/config.yaml

@@ -1,6 +1,6 @@
 defaults:
   - model : Demucs
-  - dataset : Vctk
+  - dataset : MS-SNSD
   - optimizer : Adam
   - hyperparameters : default
   - trainer : default

recipes/MS-SNSD/MS-SNSD-20/Demucs/train_config/dataset/MS-SNSD.yaml

@@ -1,10 +1,11 @@
 _target_: mayavoz.data.dataset.MayaDataset
+name : MS-SDSD
 root_dir : /Users/shahules/Myprojects/MS-SNSD
-name : dns-2020
+duration : 5
-duration : 2.0
+stride : 1
 sampling_rate: 16000
 batch_size: 32
-valid_size: 0.05
+min_valid_minutes: 25
 files:
   train_clean : CleanSpeech_training
   test_clean : CleanSpeech_training

recipes/MS-SNSD/MS-SNSD-20/Demucs/train_config/hyperparameters/default.yaml

@@ -0,0 +1,7 @@
+loss : mae
+metric : [stoi,pesq]
+lr : 0.0003
+ReduceLr_patience : 10
+ReduceLr_factor : 0.5
+min_lr : 0.000001
+EarlyStopping_factor : 10

recipes/MS-SNSD/MS-SNSD-20/Demucs/train_config/mlflow/experiment.yaml

@@ -0,0 +1,2 @@
+experiment_name : shahules/mayavoz
+run_name : demucs-ms-snsd

recipes/MS-SNSD/MS-SNSD-20/README.md

@@ -0,0 +1,17 @@
+### Microsoft Scalable Noisy Speech Dataset (MS-SNSD)
+
+ MS-SNSD is a speech datasetthat can scale to arbitrary sizes depending on the number of speakers, noise types, and Speech to Noise Ratio (SNR) levels desired.
+
+### Dataset download & setup
+- Follow steps in the official repo [here](https://github.com/microsoft/MS-SNSD) to download and setup the dataset.
+
+**References**
+```BibTex
+@article{reddy2019scalable,
+  title={A Scalable Noisy Speech Dataset and Online Subjective Test Framework},
+  author={Reddy, Chandan KA and Beyrami, Ebrahim and Pool, Jamie and Cutler, Ross and Srinivasan, Sriram and Gehrke, Johannes},
+  journal={Proc. Interspeech 2019},
+  pages={1816--1820},
+  year={2019}
+}
+```

recipes/Valentini-dataset/28spk/cli/train_config/hyperparameters/default.yaml

@@ -1,7 +0,0 @@
-loss : mae
-metric : [stoi,pesq,si-sdr]
-lr : 0.0003
-ReduceLr_patience : 5
-ReduceLr_factor : 0.2
-min_lr : 0.000001
-EarlyStopping_factor : 10

recipes/Valentini-dataset/28spk/cli/train_config/mlflow/experiment.yaml

@@ -1,2 +0,0 @@
-experiment_name : shahules/mayavoz
-run_name : Demucs + Vtck with stride + augmentations

recipes/Valentini-dataset/28spk/cli/train_config/model/DCCRN.yaml

@@ -1,25 +0,0 @@
-_target_: mayavoz.models.dccrn.DCCRN
-num_channels: 1
-sampling_rate : 16000
-complex_lstm : True
-complex_norm : True
-complex_relu : True
-masking_mode : True
-
-encoder_decoder:
-  initial_output_channels : 32
-  depth : 6
-  kernel_size : 5
-  growth_factor : 2
-  stride : 2
-  padding : 2
-  output_padding : 1
-
-lstm:
-  num_layers : 2
-  hidden_size : 256
-
-stft:
-  window_len : 400
-  hop_size : 100
-  nfft : 512

recipes/Valentini-dataset/28spk/cli/train_config/model/Demucs.yaml

@@ -1,16 +0,0 @@
-_target_: mayavoz.models.demucs.Demucs
-num_channels: 1
-resample: 4
-sampling_rate : 16000
-
-encoder_decoder:
-  depth: 4
-  initial_output_channels: 64
-  kernel_size: 8
-  stride: 4
-  growth_factor: 2
-  glu: True
-
-lstm:
-  bidirectional: False
-  num_layers: 2

recipes/Valentini-dataset/28spk/cli/train_config/model/WaveUnet.yaml

@@ -1,5 +0,0 @@
-_target_: mayavoz.models.waveunet.WaveUnet
-num_channels : 1
-depth : 9
-initial_output_channels: 24
-sampling_rate : 16000

recipes/Valentini-dataset/README.md

@@ -4,7 +4,7 @@ Clean and noisy parallel speech database. The database was designed to train and
 
 **References**
 ```BibTex
-@misc{DNSChallenge2020,
+@misc{
 title={Noisy speech database for training speech enhancement algorithms and TTS models},
 author={Valentini-Botinhao, Cassia}, year={2017},
 doi=https://doi.org/10.7488/ds/2117,

tests/utils_test.py

@@ -36,7 +36,7 @@ def test_fileprocessor_vctk():
     assert len(matching_dict) == 2
 
 
-@pytest.mark.parametrize("dataset_name", ["vctk", "dns-2020"])
+@pytest.mark.parametrize("dataset_name", ["vctk", "MS-SNSD"])
 def test_fileprocessor_names(dataset_name):
     fp = Fileprocessor.from_name(dataset_name, "clean_dir", "noisy_dir")
     assert hasattr(fp.matching_function, "__call__")