dec duration

:

enhancer/cli/train.py

@@ -60,9 +60,9 @@ def main(config: DictConfig):
 
     if parameters.get("Early_stop", False):
         early_stopping = EarlyStopping(
-            monitor="val_loss",
+            monitor=f"valid_{parameters.get('EarlyStopping_metric','loss')}",
             mode=direction,
-            min_delta=0.0,
+            min_delta=parameters.get("EarlyStopping_delta", 0.00),
             patience=parameters.get("EarlyStopping_patience", 10),
             strict=True,
             verbose=False,
@@ -93,7 +93,7 @@ def main(config: DictConfig):
 
     trainer = instantiate(config.trainer, logger=logger, callbacks=callbacks)
     trainer.fit(model)
-    trainer.test(model)
+    trainer.test(ckpt_path="best")
 
     logger.experiment.log_artifact(
         logger.run_id, f"{trainer.default_root_dir}/config_log.yaml"

enhancer/cli/train_config/config.yaml

@@ -1,6 +1,6 @@
 defaults:
-  - model : Demucs
+  - model : DCCRN
-  - dataset : Vctk
+  - dataset : DNS-2020
   - optimizer : Adam
   - hyperparameters : default
   - trainer : default

enhancer/cli/train_config/dataset/DNS-2020.yaml

@@ -1,12 +1,13 @@
 _target_: enhancer.data.dataset.EnhancerDataset
-root_dir : /Users/shahules/Myprojects/MS-SNSD
+root_dir : /scratch/c.sistc3/MS-SNSD/DNS20
 name : dns-2020
-duration : 2.0
+duration : 1.5
+stride : 1
 sampling_rate: 16000
 batch_size: 32
-valid_size: 0.05
+min_valid_minutes: 25.0
 files:
   train_clean : CleanSpeech_training
-  test_clean : CleanSpeech_training
+  test_clean : CleanSpeech_testing
   train_noisy : NoisySpeech_training
-  test_noisy : NoisySpeech_training
+  test_noisy : NoisySpeech_testing

enhancer/cli/train_config/dataset/Vctk.yaml

@@ -1,11 +1,11 @@
 _target_: enhancer.data.dataset.EnhancerDataset
 name : vctk
 root_dir : /scratch/c.sistc3/DS_10283_2791
-duration : 4.5
+duration : 2
-stride : 2
+stride : 1
 sampling_rate: 16000
 batch_size: 32
-valid_minutes : 15
+min_valid_minutes : 25
 files:
   train_clean : clean_trainset_28spk_wav
   test_clean : clean_testset_wav

enhancer/cli/train_config/hyperparameters/default.yaml

@@ -1,7 +1,8 @@
-loss : mae
+loss : si-snr
-metric : [stoi,pesq,si-sdr]
+metric : [stoi,pesq]
-lr : 0.0003
+lr : 0.001
-ReduceLr_patience : 5
+ReduceLr_patience : 10
-ReduceLr_factor : 0.2
+Early_stop : False
-min_lr : 0.000001
+ReduceLr_factor : 0.5
-EarlyStopping_factor : 10
+min_lr : 0.0000001
+EarlyStopping_patience : 10

enhancer/cli/train_config/mlflow/experiment.yaml

@@ -1,2 +1,2 @@
 experiment_name : shahules/enhancer
-run_name : Demucs + Vtck with stride + augmentations
+run_name : dccrn-dns20

enhancer/cli/train_config/model/DCCRN.yaml

@@ -4,7 +4,7 @@ sampling_rate : 16000
 complex_lstm : True
 complex_norm : True
 complex_relu : True
-masking_mode : True
+masking_mode : "E"
 
 encoder_decoder:
   initial_output_channels : 32

enhancer/cli/train_config/model/Demucs.yaml

@@ -12,5 +12,5 @@ encoder_decoder:
   glu: True
 
 lstm:
-  bidirectional: False
+  bidirectional: True
   num_layers: 2

enhancer/cli/train_config/trainer/default.yaml

@@ -2,14 +2,14 @@ _target_: pytorch_lightning.Trainer
 accelerator: gpu
 accumulate_grad_batches: 1
 amp_backend: native
-auto_lr_find: True
+auto_lr_find: False
 auto_scale_batch_size: False
 auto_select_gpus: True
 benchmark: False
 check_val_every_n_epoch: 1
 detect_anomaly: False
 deterministic: False
-devices: 2
+devices: 1
 enable_checkpointing: True
 enable_model_summary: True
 enable_progress_bar: True
@@ -23,7 +23,7 @@ limit_test_batches: 1.0
 limit_train_batches: 1.0
 limit_val_batches: 1.0
 log_every_n_steps: 50
-max_epochs: 200
+max_epochs: 100
 max_steps: -1
 max_time: null
 min_epochs: 1
@@ -38,9 +38,8 @@ precision: 32
 profiler: null
 reload_dataloaders_every_n_epochs: 0
 replace_sampler_ddp: True
-strategy: ddp
+strategy: null
 sync_batchnorm: False
 tpu_cores: null
 track_grad_norm: -1
 val_check_interval: 1.0
-weights_save_path: null

enhancer/data/dataset.py

@@ -135,7 +135,7 @@ class TaskDataset(pl.LightningDataModule):
             speaker_index = rng.choice(possible_indices)
             possible_indices.remove(speaker_index)
             speaker_name = all_speakers[speaker_index]
-            print(f"Selected f{speaker_name} for valid")
+            print(f"Selected {speaker_name} for valid")
             file_indices = [
                 i
                 for i, file in enumerate(data)

enhancer/loss.py

@@ -192,7 +192,7 @@ class Si_snr(nn.Module):
         super().__init__()
 
         self.loss_fun = ScaleInvariantSignalNoiseRatio(**kwargs)
-        self.higher_better = True
+        self.higher_better = False
         self.name = "si_snr"
 
     def forward(self, prediction: torch.Tensor, target: torch.Tensor):
@@ -203,7 +203,7 @@ class Si_snr(nn.Module):
                     got {prediction.size()} and {target.size()} instead"""
             )
 
-        return self.loss_fun(prediction, target)
+        return -1 * self.loss_fun(prediction, target)
 
 
 LOSS_MAP = {

enhancer/utils/transforms.py

@@ -85,7 +85,7 @@ class ConviSTFT(ConvFFT):
             input = torch.cat([real, imag], 1)
         out = F.conv_transpose1d(input, self.weight, stride=self.hop_size)
         coeff = self.window.unsqueeze(1).repeat(1, 1, input.size(-1)) ** 2
-        coeff.to(input.device)
+        coeff = coeff.to(input.device)
         coeff = F.conv_transpose1d(coeff, self.enframe, stride=self.hop_size)
         out = out / (coeff + 1e-8)
         pad = self.window_len - self.hop_size

hpc_entrypoint.sh

@@ -34,6 +34,8 @@ pwd
 
 #python transcriber/tasks/embeddings/timit.py --directory /scratch/$USER/TIMIT/data/lisa/data/timit/raw/TIMIT/TRAIN --output ./data/train
 #python transcriber/tasks/embeddings/timit.py --directory /scratch/$USER/TIMIT/data/lisa/data/timit/raw/TIMIT/TEST --output ./data/test
+# mv /scratch/c.sistc3/MS-SNSD/DNS20/CleanSpeech_testing /scratch/c.sistc3/MS-SNSD/DNS30/CleanSpeech_testing
+# mv /scratch/c.sistc3/MS-SNSD/DNS20/NoisySpeech_testing /scratch/c.sistc3/MS-SNSD/DNS30/NoisySpeech_testing
 
 echo "Start Training..."
 python enhancer/cli/train.py