diff --git a/vall_e/config.py b/vall_e/config.py index 5d8c1b7..6751977 100755 --- a/vall_e/config.py +++ b/vall_e/config.py @@ -198,11 +198,7 @@ class Model: return 256 if self.size == "half": return 512 - if self.size == "full": - return 1024 - if self.size == "double": - return 2048 - raise ValueError + return 1024 @property def heads(self): @@ -216,17 +212,15 @@ class Model: return 4 if self.size == "half": return 8 - if self.size == "full": - return 16 - if self.size == "double": - return 32 - raise ValueError + return 16 @property def layers(self): if isinstance(self.size, dict) and hasattr(self.size, "layers"): return self.size['layers'] + if self.size == "double": + return 24 return 12 @dataclass() diff --git a/vall_e/engines/base.py b/vall_e/engines/base.py index 721e724..7545b3c 100755 --- a/vall_e/engines/base.py +++ b/vall_e/engines/base.py @@ -77,6 +77,8 @@ class Engine(): @property def _training(self): + if not hasattr(self, "_cfg"): + return True return self._cfg.training @property @@ -199,7 +201,6 @@ class Engine(): with torch.autocast("cuda", dtype=cfg.trainer.dtype, enabled=cfg.trainer.amp): self.forward(*args, **kwargs) losses = self.gather_attribute("loss") - print(self.module.loss) loss = torch.stack([*losses.values()]).sum() stats = {} diff --git a/vall_e/models/ar.py b/vall_e/models/ar.py index bd5dad2..dcf1e72 100755 --- a/vall_e/models/ar.py +++ b/vall_e/models/ar.py @@ -132,14 +132,14 @@ def example_usage(): from tqdm import tqdm device = "cuda" - x8 = partial(repeat, pattern="t -> t l", l=2) + x8 = partial(repeat, pattern="t -> t l", l=cfg.models.prom_levels) symmap = {'': 1, '': 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} def tokenize(content, lang_marker="en"): split = content.split(" ") phones = [f""] + [ " " if not p else p for p in split ] + [f""] return torch.tensor([*map(symmap.get, phones)]).to() - qnt = torch.load("data/qnt.pt")[0].t()[:, :2].to(device) + qnt = torch.load("data/qnt.pt")[0].t()[:, :cfg.models.prom_levels].to(device) text_list = [ #torch.tensor([1, 2, 3], device=device), @@ -179,7 +179,7 @@ def example_usage(): stats = {"step": i} stats |= engine.traverse(text_list=text_list, proms_list=proms_list, resps_list=resps_list) - t.set_description(f"{stats}") + tqdm.write(f"{stats}") sample("init", 75) train() diff --git a/vall_e/models/base.py b/vall_e/models/base.py index a02d325..c74bace 100755 --- a/vall_e/models/base.py +++ b/vall_e/models/base.py @@ -398,7 +398,7 @@ def example_usage(): from .nar import NAR device = "cuda" - x8 = partial(repeat, pattern="t -> t l", l=2) + x8 = partial(repeat, pattern="t -> t l", l=cfg.models.prom_levels) symmap = {'': 1, '': 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} def tokenize(content, lang_marker="en"): split = content.split(" ") @@ -409,15 +409,18 @@ def example_usage(): 'n_tokens': 1024, 'd_model': 1024, 'n_heads': 16, - 'n_layers': 24, + 'n_layers': 12, } models = { "ar": AR(**kwargs).to(device), "nar": NAR(**kwargs).to(device) } + + for name, model in models.items(): + print(f"{name} parameter count: {sum(p.numel() for p in model.parameters() if p.requires_grad)}") + engines = Engines({ name: Engine(model=model, optimizer=torch.optim.AdamW(model.parameters(), lr=1e-4)) for name, model in models.items() }) train = True - - qnt = torch.load("data/qnt.pt")[0].t()[:, :2].to(device) + qnt = torch.load("data/qnt.pt")[0].t()[:, :cfg.models.prom_levels].to(device) text_list = [ tokenize("ˈ a ɪ w ɪ l nˌ ɑː t ˈ æ s k ɐ sˈ ɛ k ə n d tˈ a ɪ m").to(device), #tokenize("ˌ ɔ n ɡˌ o ʊ ɪ ŋ hˈ o ʊ m ð ə tˈ uː f ɹˈ ɛ n d z fˈ a ʊ n d ɐ lˈ ɛ ɾ ɚ f ɹ ʌ m ˈ æ θ o ʊ z , hˌ uː d ɪ zˈ a ɪ ɚ d ðˌ ɛ m t ə mˈ iː t hˌ ɪ m æ t ð ə ɡ ɹˈ æ n d t ʃˈ ɑː ɹ l ɪ mˌ æ ɡ n i ɔ n ð ə fˈ ɑː l o ʊ ɪ ŋ dˈ e ɪ .").to(device), @@ -455,12 +458,12 @@ def example_usage(): t = trange(60) for i in t: stats = {"step": i} + """ for name, engine in engines.items(): stats |= engine.traverse(text_list=text_list, proms_list=proms_list, resps_list=resps_list) """ stats = engines.step({"text_list": text_list, "proms_list": proms_list, "resps_list": resps_list}) - """ - t.set_description(f"{stats}") + tqdm.write(f"{stats}") else: for name, engine in engines.items(): engine.module.load_state_dict(torch.load(f"./data/{name}.pth")) diff --git a/vall_e/models/nar.py b/vall_e/models/nar.py index e33e2b3..aed1bf9 100755 --- a/vall_e/models/nar.py +++ b/vall_e/models/nar.py @@ -144,7 +144,7 @@ def example_usage(): return torch.tensor([*map(symmap.get, phones)]).to() # to-do: unmangle this and the resp shit - qnt = torch.load("data/qnt.pt")[0].t()[:, :2].to(device) + qnt = torch.load("data/qnt.pt")[0].t()[:, :cfg.models.prom_levels].to(device) text_list = [ #torch.tensor([1, 2, 3], device=device), @@ -180,7 +180,7 @@ def example_usage(): stats = {"step": i} stats |= engine.traverse(text_list=text_list, proms_list=proms_list, resps_list=resps_list) - t.set_description(f"{stats}") + tqdm.write(f"{stats}") sample("init") train()