draft

eval.py

+import museval
+import argparse
+
+if __name__ == '__main__':
+    # Training settings
+    parser = argparse.ArgumentParser(
+        description='MUSDB18 Evaluation',
+        add_help=False
+    )
+
+    parser.add_argument(
+        '--eval-path',
+        type=str,
+        help='Path evaluation root folder'
+    )
+
+    args = parser.parse_args()
+    method = museval.EvalStore()
+    method.add_eval_dir(args.eval_path)
+    print(method)

test_demucs.py

@@ -12,7 +12,7 @@ def separate(
     audio,
     rate,
     model,
-    shifts=5,
+    shifts=1,
     device="cpu"
 ):
 

test_unet.py

@@ -4,6 +4,7 @@ import norbert
 import numpy as np
 import soundfile as sf
 import torch
+from tqdm import tqdm
 # from torch_specinv import griffin_lim
 
 from evaluator.music_demixing import MusicDemixingPredictor
@@ -47,7 +48,7 @@ def istft(X, n_fft=4096, n_hopsize=1024):
 def separate(
     audio,
     x_umx_target,
-    niter=0,
+    niter=1,
     softmask=False,
     alpha=1.0,
     residual_model=False,
@@ -90,7 +91,11 @@ def separate(
 
     with torch.no_grad():
         mag = X.abs()
-        masked_tf_rep = x_umx_target(mag.unsqueeze(0)).squeeze() * mag
+        frames = 2584
+        masked_tf_rep = []
+        for chunk in tqdm(mag.split(frames, dim=-1)):
+            masked_tf_rep += [x_umx_target(chunk.unsqueeze(0)).squeeze()]
+        masked_tf_rep = torch.cat(masked_tf_rep, -1) * mag
 
     Vj = masked_tf_rep.cpu().numpy()
     if softmask: