cleaning

nnet/xsets.py

@@ -603,16 +603,10 @@ class IdMapSet(Dataset):
             stop = int(self.idmap.stop[index])
             # add this in case the segment is too short
             if stop - start <= self.min_duration * nfo.samplerate:
-                print("Segment trop court")
-                print(f"Get: {stop - start} and need {self.min_duration * nfo.samplerate}")
-                print(f"self.min_duration = {self.min_duration}, nfo.samplerate = {nfo.samplerate}")
-                print(f"start = {start}, stop = {stop}\t start//160 {start//160}, stop {stop//160}")
                 middle = start + (stop - start) // 2
-                print(f"middle = {middle}")
                 start = max(0, int(middle - (self.min_duration * nfo.samplerate / 2)))
 
                 stop = int(start + self.min_duration * nfo.samplerate)
-                print(f"Apres modification: start = {start}, stop = {stop} total duration: {stop - start}")
             sig, _ = soundfile.read(f"{self.data_root_path}/{self.idmap.rightids[index]}.{self.file_extension}",
                                     start=start,
                                     stop=stop)
@@ -621,7 +615,6 @@ class IdMapSet(Dataset):
         if self.transform_pipeline is not None:
             sig, _, ___, _____, _t, _s = self.transforms((sig, 0,  0, 0, 0, 0))
 
-
         return torch.from_numpy(sig).type(torch.FloatTensor), \
                self.idmap.leftids[index], \
                self.idmap.rightids[index], \

nnet/xvector.py

@@ -956,7 +956,7 @@ def extract_embeddings(idmap_name,
                        file_extension=file_extension,
                        transform_pipeline=transform_pipeline,
                        frame_rate=int(1. / frame_shift),
-                       min_duration=(model.context_size() + 2) * frame_shift
+                       min_duration=(model.context_size() + 2) * frame_shift * 2
                        )
 
     dataloader = DataLoader(dataset,