backwards compat for my shitty old weights (was testing if disabling AudioEmbedding summing magically made things better (it did not))

vall_e/config.py

@@ -619,6 +619,31 @@ class Config(_Config):
 	#if not isinstance(self.bitsandbytes, type):
 		self.bitsandbytes = BitsAndBytes(**self.bitsandbytes)
 
+# Preserves the old behavior
+class NaiveTokenizer:
+	def get_vocab( self ):
+		"""
+		if cfg.dataset.use_hdf5 and 'symmap' in cfg.hdf5:
+			return json.loads( cfg.hdf5['symmap'].asstr()[()] )
+		"""
+		return {'<s>': 1, '</s>': 2, ' ': 3, '.': 4, ',': 5, '!': 6, '?': 7, 'p': 7, 'iː': 8, 'ɚ': 9, 'ˌ': 10, 'dˌ': 11, 'mˌ': 12, 'd': 13, 'ɹ': 14, 'tˈ': 15, 'pˌ': 16, 'uː': 17, 'l': 18, 'æ': 19, 'ɛ': 20, 'ɪ': 21, 'j': 22, 'ʊ': 23, 't': 24, 'n': 25, 'v': 26, 'a': 27, 'o': 28, 'ŋ': 29, 'w': 30, 'ʌ': 31, 'hˈ': 32, 'ɡˈ': 33, 'ə': 34, 'θˈ': 35, 'dˈ': 36, 'wˌ': 37, 'h': 38, 'z': 39, 'k': 40, 'ð': 41, 'ɡˌ': 42, 'ˈ': 43, 'fˈ': 44, 'i': 45, 's': 46, 'ʃ': 47, 'wˈ': 48, 'ðˈ': 49, 'ɹˈ': 50, 'lˈ': 51, 'ɡ': 52, 'oː': 53, 'mˈ': 54, 'e': 55, 'ɑː': 56, 'nˈ': 57, 'm': 58, 'θˌ': 59, 'sˈ': 60, 'f': 61, 'ɔː': 62, 'hˌ': 63, 'b': 64, 'jˈ': 65, 'ɐ': 66, 'ʒˈ': 67, 'θ': 68, 'bˈ': 69, 'ɾ': 70, 'ɜː': 71, 'ʌˈ': 72, 'ʃˌ': 73, 'bˌ': 74, 'kˈ': 75, 'ɔ': 76, 'zˈ': 77, 'ᵻ': 78, 'kˌ': 79, 'vˈ': 80, 'fˌ': 81, 'ʒ': 82, 'ʃˈ': 83, 'ɹˌ': 84, 'tˌ': 85, 'pˈ': 86, 'ðˌ': 87, 'sˌ': 88, 'nˌ': 89, 'lˌ': 90, '̩': 91, 'ʔ': 92, 'vˌ': 93, 'ɪˈ': 94, '"': 95, 'ɪˌ': 96, 'ʒˌ': 97, 'uːˌ': 98, 'ʊˈ': 99, 'jˌ': 100, 'uːˈ': 101, 'iːˈ': 102, 'zˌ': 103, '.ˈ': 104, '…': 105, 'ŋˌ': 106, 'ɐˌ': 107, '—ˈ': 108, 'iˌ': 109, 'iːˌ': 110, 'ɛː': 111, ')': 112, ')ˈ': 113, '(': 114, 'u': 115, '-': 116, 'ɖˈ': 117, 'iˈ': 118, 'ʰˈ': 119, 'ɟˈ': 120, '̃': 121, 'eː': 122, 'ɾˈ': 123, 'r': 124, 'ʰ': 125, '-ˌ': 126, 'ɫ': 127, 'q': 128, '—': 129, 'ʊˌ': 130, 'aː': 131, 'cˈ': 132, '…ˈ': 133, 'c': 134, 'ɳ': 135, 'ɐˈ': 136, 'x': 137, 'ʔˌ': 138, '.ˌ': 139, 'ɑ': 140, '?ˈ': 141, '̩ˈ': 142, '"ˈ': 143, ',ˈ': 144, 'ŋˈ': 145, 'əˌ': 146, '!ˈ': 147, '"ˌ': 148, '?ˌ': 149, ',ˌ': 150, '—ˌ': 151, '̩ˌ': 152, 'əˈ': 153, '!ˌ': 154, 'ɬ': 155, 'ʲ': 156, '¡': 157, 'ɯ': 158, 'qˌ': 159, 'ʑ': 160, 'ʑˈ': 161, '¿': 162, 'ɑːˈ': 163, 'iːː': 164, 'ɛˈ': 165, '¡ˈ': 166, 'æˈ': 167, 'ç': 168, 'ɾˌ': 169, 'ᵻˈ': 170, 'xˈ': 171, 'ɔːˈ': 172, ';': 173, 'ɬˌ': 174, ':': 175, 'ʔˈ': 176, 'ɑːˌ': 177, 'ɬˈ': 178, '”': 179, '“': 180, '“ˈ': 181, '“ˌ': 182, ';ˈ': 183, ';ˌ': 184, ':ˈ': 185, '1': 186, 'rˈ': 187, 'qˈ': 188, 'ᵻˌ': 189, 'ä': 190, '̞ˌ': 191, '̞': 192, 'ũˌ': 193, 'ʑˌ': 194, 'ᵝ': 195, 'ɽ': 196, 'ʲˌ': 197, 'ᵝˌ': 198, 'ũ': 199, 'ũˈ': 200, 'äˌ': 201, 'ɕ': 202, 'ɕˌ': 203, 'ɽˌ': 204, 'çˌ': 205, '…ˌ': 206, '̞ˈ': 207, 'äˈ': 208, 'ɽˈ': 209, 'ɸˌ': 210, 'ɴ': 211, 'ɸˈ': 212, 'ɕˈ': 213, 'ɸ': 214, 'ᵝˈ': 215, 'ʲˈ': 216, 'ĩ': 217, 'çˈ': 218, 'ĩˌ': 219, 'oˌ': 220, 'eˈ': 221, 'ʍ': 222, 'eˌ': 223, 'uˌ': 224, 'ʍˌ': 225, 'uˈ': 226, 'oˈ': 227, 'aˈ': 228}
+
+	def encode( self, s ):
+		symmap = self.get_vocab()
+		phones = " ".join( list(s) )
+
+		# do merge
+		for merge in [ "\u02C8", "\u02CC", "\u02D0" ]:
+			phones = phones.replace( f' {merge}', merge )
+
+		phones = phones.split(" ")
+		# cleanup
+		phones = [ p for i, p in enumerate(phones) if p not in [" "] or ( p in [" "] and p != phones[i-1] ) ]
+		# add bos / eos
+		phones = ["<s>"] + [ " " if not p else p for p in phones ] + ["</s>"]
+		# tokenize
+		return [*map(symmap.get, phones)]
+
 
 cfg = Config.from_cli()
 
@@ -636,6 +661,7 @@ try:
 	cfg.tokenizer = (cfg.relpath if cfg.cfg_path is not None else Path("./data/")) / cfg.tokenizer
 	cfg.tokenizer = PreTrainedTokenizerFast(tokenizer_file=str(cfg.tokenizer))
 except Exception as e:
+	cfg.tokenizer = NaiveTokenizer()
 	print("Error while parsing tokenizer:", e)
 	pass
 

vall_e/data.py

@@ -33,8 +33,7 @@ def get_phone_symmap():
 	return cfg.tokenizer.get_vocab()
 
 def tokenize( phones ):
-	return tokenizer.encode( "".join(phones) )
+	return cfg.tokenizer.encode( "".join(phones) )
-	#return [*map(get_phone_symmap.get, _get_phones(path))]
 
 def get_lang_symmap():
 	return {

vall_e/inference.py

@@ -13,7 +13,7 @@ from .utils import to_device
 from .config import cfg
 from .models import get_models
 from .engines import load_engines, deepspeed_available
-from .data import get_phone_symmap, get_lang_symmap, _load_quants, _cleanup_phones
+from .data import get_phone_symmap, get_lang_symmap, _load_quants, _cleanup_phones, tokenize
 
 if deepspeed_available:
 	import deepspeed
@@ -91,8 +91,9 @@ class TTS():
 			return text
 
 		content = g2p.encode(text, language=language)
+		tokens = tokenize( content )
 
-		return torch.tensor(cfg.tokenizer.encode( "".join( content ) ))
+		return torch.tensor( tokens )
 
 	def encode_lang( self, language ):
 		symmap = get_lang_symmap()

vall_e/models/ar_nar.py

@@ -350,7 +350,7 @@ def example_usage():
 		tokenize("ˈaɪ wɪl nˌɑːt ˈæsk ɐ sˈɛkənd tˈaɪm").to(device),
 	]
 	proms_list = [
-		qnt.to(device),
+		qnt[:75, :].to(device),
 	]
 	resps_list = [
 		qnt.to(device),
@@ -369,6 +369,8 @@ def example_usage():
 		'n_experts': 1,
 
 		'l_padding': 8 if cfg.fp8.enabled else 0,
+
+		'config': cfg.model
 	}
 	"""
 	kwargs = {
@@ -388,7 +390,7 @@ def example_usage():
 	"""
 
 	model = AR_NAR(**kwargs).to(device)
-	steps = 750
+	steps = 500
 	optimizer = ml.Prodigy(model.parameters(), lr=1.0)
 	#optimizer = ml.Adagrad(model.parameters(), lr=1.0e-2)
 	#optimizer = ml.AdamW(model.parameters(), lr=1.0e-4)

vall_e/models/base.py

@@ -365,13 +365,13 @@ class Base(nn.Module):
 			self.proms_emb = AudioEmbedding(
 				[n_prom_tokens] * self.n_prom_levels, d_model,
 				levels=self.n_prom_levels if self.version > 3 else None,
-				sums=self.config.audio_embedding_sums
+				sums=self.config.audio_embedding_sums if self.config is not None else True,
 			)
 			# [1025] + [1024] * 8
 			self.resps_emb = AudioEmbedding(
 				[n_resp_tokens] + [n_resp_tokens - 1] * (self.n_resp_levels - 1), d_model,
 				levels=self.n_resp_levels if self.version > 3 else None,
-				sums=self.config.audio_embedding_sums
+				sums=self.config.audio_embedding_sums if self.config is not None else True
 			)