use least squares weights

lstsq.py

@@ -182,7 +182,7 @@ if __name__ == "__main__":
             x = stems[0, i:i+valid_length]
             y = stems[1:, i:i+valid_length]
 
-            pred1 = unet_separate(x, unet, device=device)
+            pred1 = unet_separate(x, unet, device=device, niter=1)
             pred2 = demucs_separate(x, 44100, demucs, shifts=3, device=device)
             min_length = min(pred1.shape[-1], pred2.shape[-1])
             pred1 = pred1[..., :min_length]

test_fuser.py

@@ -10,8 +10,8 @@ from evaluator.music_demixing import MusicDemixingPredictor
 
 
 fusion_weights = np.array([0.27203053, 0.73882418])
-fusion_weights_batch = np.array([[0.25, 0.27, 0.5, 0.6],
-                                 [0.75, 0.73, 0.5, 0.4]])
+fusion_weights_batch = np.array([[0.12199835471686171, 0.051671789064391405, 0.35361103076576583, 0.4857059599263529],
+                                 [0.8812474887611046, 0.9477684070510808, 0.6320408659140799, 0.5168781960641784]])
 fusion_weights_full = np.array(
     [[0.12421362233786719, 0.053170121998740195, -0.023004887368765208, -0.030820051344791908, 0.7362167118453092, -0.2011487819137626, -0.1254105898553318, -0.11682101700727994, 0.14540460690976179],
      [0.06420106209073237, 0.07177930713127706, 0.0035666737989309845, 0.0005502465774743552, -