diff --git a/vall_e/config.py b/vall_e/config.py index 54c24e6..cb144b7 100755 --- a/vall_e/config.py +++ b/vall_e/config.py @@ -140,6 +140,8 @@ class Dataset: tasks_list: list[str] = field(default_factory=lambda: ["tts"]) + continuous: bool = False # VALL-E continuous, as explained in the paper + @property def min_phones(self): return self.phones_range[0] @@ -156,20 +158,13 @@ class Dataset: @dataclass() class Model: name: str = "" - size: str = "full" + size: str | float | dict = "full" resp_levels: int = 1 prom_levels: int = 8 tasks: int = 8 # ["tts", "ns", "sr", "tse", "cse", "nse"] and leaves two more for anything else I want (like "svc") arch_type: str = "transformer" training: bool = True - @property - def scale(self): - if self.size == "quarter": - return 0.25 - if self.size == "half": - return 0.5 - return 1.0 @property def full_name(self): @@ -187,30 +182,52 @@ class Model: @property def tokens(self): + if isinstance(self.size, dict) and hasattr(self.size, "tokens"): + return self.size['tokens'] + return 1024 @property def dim(self): + if isinstance(self.size, dict) and hasattr(self.size, "dim"): + return self.size['dim'] + + if isinstance(self.size, float): + return math.floor(1024 * self.size) + if self.size == "quarter": return 256 if self.size == "half": return 512 if self.size == "full": return 1024 + if self.size == "double": + return 2048 raise ValueError @property def heads(self): + if isinstance(self.size, dict) and hasattr(self.size, "heads"): + return self.size['heads'] + + if isinstance(self.size, float): + return math.floor(16 * self.size) + if self.size == "quarter": return 4 if self.size == "half": return 8 if self.size == "full": return 16 + if self.size == "double": + return 32 raise ValueError @property def layers(self): + if isinstance(self.size, dict) and hasattr(self.size, "layers"): + return self.size['layers'] + return 12 @dataclass() diff --git a/vall_e/data.py b/vall_e/data.py index 1153427..b217fee 100755 --- a/vall_e/data.py +++ b/vall_e/data.py @@ -313,7 +313,15 @@ class Dataset(_Dataset): noise_scale = 0.25 # text-to-speech if task == "tts": - proms = self.sample_prompts(spkr_name, ignore=path) if random.random() < cfg.dataset.random_utterance else resps + trim_length = int(cfg.dataset.prompt_duration * 75) + # VALL-E continuous + # ignore if target utterance is shorter than prompt duration + # to-do: actually do this for the AR only as I don't think the paper trained the NAR for this + if cfg.dataset.continuous and trim_length > resps.shape[0]: + proms = resps[:trim_length, :] + resps = resps[trim_length:, :] + else: + proms = self.sample_prompts(spkr_name, ignore=path) if random.random() < cfg.dataset.random_utterance else resps # noise suppression || speech removal elif task == "ns" or task == "sr": # sample random noise diff --git a/vall_e/emb/qnt.py b/vall_e/emb/qnt.py index 03f9983..7a2f9dc 100755 --- a/vall_e/emb/qnt.py +++ b/vall_e/emb/qnt.py @@ -185,11 +185,18 @@ def encode_from_file(path, device="cuda"): # trims from the start, up to `target` def trim( qnt, target ): length = max( qnt.shape[0], qnt.shape[1] ) - start = 0 - end = start + target - if end >= length: - start = length - target - end = length + if target > 0: + start = 0 + end = start + target + if end >= length: + start = length - target + end = length + # negative length specified, trim from end + else: + start = length + target + end = length + if start < 0: + start = 0 return qnt[start:end] if qnt.shape[0] > qnt.shape[1] else qnt[:, start:end] diff --git a/vall_e/models/ar.py b/vall_e/models/ar.py index c8ac130..ec825ab 100755 --- a/vall_e/models/ar.py +++ b/vall_e/models/ar.py @@ -23,6 +23,8 @@ class AR(Base): @property def arch_type(self) -> bool: + if hasattr(self, "_cfg"): + return self._cfg.arch_type return cfg.models.ar.arch_type @property @@ -31,6 +33,8 @@ class AR(Base): @property def n_resp_levels(self) -> int: + if hasattr(self, "_cfg"): + return self._cfg.resp_levels return cfg.models.ar.resp_levels @property diff --git a/vall_e/models/base.py b/vall_e/models/base.py index 3b3f7ff..c529ac7 100755 --- a/vall_e/models/base.py +++ b/vall_e/models/base.py @@ -387,6 +387,9 @@ def example_usage(): from .ar import AR from .nar import NAR + from ..models import get_models + from ..config import Model as ModelCfg + device = "cuda" x8 = partial(repeat, pattern="t -> t l", l=2) 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} @@ -395,13 +398,20 @@ def example_usage(): phones = [f""] + [ " " if not p else p for p in split ] + [f""] return torch.tensor([*map(symmap.get, phones)]).to() + models = get_models({ + "ar": Model(name="ar", resp_levels=1, prom_levels=8, tasks=8, training=True, size=1), + "nar": Model(name="nar", resp_levels=7, prom_levels=8, tasks=8, training=True, size=1)} + ) + """ + model_cfg = ModelCfg() kwargs = { - 'n_tokens': 1024, - 'd_model': 1024, - 'n_heads': 16, - 'n_layers': 12, + 'n_tokens': model_cfg.tokens, + 'd_model': model_cfg.dim, + 'n_heads': model_cfg.heads, + 'n_layers': model_cfg.layers, } models = { "ar": AR(**kwargs).to(device), "nar": NAR(**kwargs).to(device) } + """ engines = Engines({ name: Engine(model=model, optimizer=torch.optim.AdamW(model.parameters(), lr=1e-4)) for name, model in models.items() }) train = True diff --git a/vall_e/models/nar.py b/vall_e/models/nar.py index 953f419..0edefe5 100755 --- a/vall_e/models/nar.py +++ b/vall_e/models/nar.py @@ -17,6 +17,8 @@ class NAR(Base): @property def arch_type(self) -> bool: + if hasattr(self, "_cfg"): + return self._cfg.arch_type return cfg.models.nar.arch_type @property @@ -29,6 +31,8 @@ class NAR(Base): @property def n_resp_levels(self) -> int: + if hasattr(self, "_cfg"): + return self._cfg.resp_levels return cfg.models.nar.resp_levels @property