ugh

scripts/process_dataset.py

@@ -144,7 +144,7 @@ for dataset_name in sorted(os.listdir(f'./{input_audio}/')):
 					i = i + 1
 
 					outpath = Path(f'./{output_dataset}/{dataset_name}/{speaker_id}/{fname}_{id}.{extension}')
-					text = metadata[filename]["text"]
+					text = segment["text"]
 
 					if len(text) == 0:
 						continue

vall_e/emb/qnt.py

@@ -285,6 +285,7 @@ def encode(wav: Tensor, sr: int = cfg.sample_rate, device="cuda", levels=cfg.mod
 	wav = convert_audio(wav, sr, model.sample_rate, model.channels)
 	wav = wav.to(device)
 
+	with torch.autocast("cuda", dtype=cfg.inference.dtype, enabled=cfg.inference.amp):
 		encoded_frames = model.encode(wav)
 	qnt = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1)  # (b q t)