more tweaks

vall_e/emb/similar.py

@@ -40,6 +40,8 @@ def load_audio( path ):
 
 	return waveform, sr
 
+tts = None
+
 def process(
 	speaker_path,
 	yaml,
@@ -52,7 +54,12 @@ def process(
 
 	verbose=False,
 	metadata_path=None,
+
+	maximum_duration=0,
+	#use_faiss=True,
 ):
+	global tts
+
 	cfg.set_audio_backend(audio_backend)
 	artifact_extension = cfg.audio_backend_extension
 
@@ -60,7 +67,8 @@ def process(
 	cfg.inference.amp = amp # False
 
 	# easy way to load the model and handle encoding audio
-	tts = init_tts( yaml=yaml, restart=False, device=device, dtype=dtype )
+	if tts is None:
+		tts = init_tts( yaml=yaml, restart=False, device=device, dtype=dtype )
 
 	queue = []
 	features = {}
@@ -69,6 +77,13 @@ def process(
 
 	mfcc = T.MFCC(sample_rate=cfg.sample_rate)
 
+	"""
+	# too slow
+	if use_faiss:
+		import faiss
+		index = None
+	"""
+
 	# compute features (embeddings if quantized already, MFCC features if raw audio)
 	for filename in tqdm(os.listdir(f'./{speaker_path}/'), desc=f"Encoding '{speaker_path}'", disable=not verbose):
 		extension = filename.split(".")[-1]
@@ -91,20 +106,40 @@ def process(
 			if lang != "en":
 				phn = phn.replace(f"({metadata['language']})", "")
 
-			features[filename] = tts.text_embedding( phn )
+			embedding = tts.text_embedding( phn )
 		else:
 			# treat embeddings as features, if provided quantized audio
 			if extension in artifact_extension:
 				artifact = np.load(f'./{speaker_path}/{filename}.{extension}', allow_pickle=True)[()]
 				qnt = torch.from_numpy(artifact["codes"].astype(int))[0].t().to(dtype=torch.int16, device=device)
-				qnt = trim( qnt, int( cfg.dataset.frames_per_second * 3 ) )
+				if maximum_duration > 0:
+					qnt = trim( qnt, int( cfg.dataset.frames_per_second * maximum_duration ) )
 				
-				features[filename] = tts.audio_embedding( qnt )
+				embedding = tts.audio_embedding( qnt )
 			# try and extract features from the raw audio itself
 			else:
 				# qnt = tts.encode_audio(f'./{speaker_path}/{filename}', trim_length=3.0).to(device)
 				wav, sr = load_audio( f'./{speaker_path}/{filename}.{extension}' )
-				features[filename] = mfcc(wav.to(device))[0].t()
+				embedding = mfcc(wav.to(device))[0].t()
+		
+		features[filename] = embedding
+		
+		"""
+		if use_faiss:
+			if index is None:
+				shape = embedding.shape
+				index = faiss.IndexFlatL2(shape[1])
+
+			index.add(embedding.cpu())
+
+		if verbose:
+			for filename, embedding in features.items():
+				D, I = index.search(embedding.cpu(), k=3)
+				# print(f'{filename}: {I[1:]}')
+
+		if metadata_path is not None:
+			index.save(metadata_path)
+		"""
 
 	keys = list(features.keys())
 	key_range = range(len(keys))
@@ -126,6 +161,10 @@ def process(
 		
 		similarities[key] = similarity
 
+		# combinations() doesn't have swapped keys
+		if swapped_key not in similarities:
+			similarities[swapped_key] = similarity
+
 		if index_a not in sorted_similarities:
 			sorted_similarities[index_a] = {}
 		if index_b not in sorted_similarities[index_a]:
@@ -176,11 +215,12 @@ def main():
 
 	parser.add_argument("--yaml", type=Path)
 	parser.add_argument("--text", action="store_true")
+	parser.add_argument("--maximum-duration", type=float, default=3.0)
 
 	parser.add_argument("--audio-backend", type=str, default="encodec")
-	parser.add_argument("--dtype", type=str, default="bfloat16")
+	parser.add_argument("--dtype", type=str, default="float16")
 	parser.add_argument("--amp", action="store_true")
-	parser.add_argument("--device", type=str, default="cuda")
+	parser.add_argument("--device", type=str, default="cpu") # unironically faster
 	
 	args = parser.parse_args()
 
@@ -199,6 +239,7 @@ def main():
 				metadata_path=cfg.metadata_dir / f'{speaker_name}.json',
 				yaml=args.yaml,
 				text=args.text,
+				maximum_duration=args.maximum_duration,
 
 				audio_backend=args.audio_backend,
 				device=args.device,
@@ -224,6 +265,7 @@ def main():
 			speaker_path=args.input_speaker,
 			yaml=args.yaml,
 			text=args.text,
+			maximum_duration=args.maximum_duration,
 
 			audio_backend=args.audio_backend,
 			device=args.device,