mrq/vall-e
5
9
Fork 0
Code Issues 8 Pull Requests Packages Projects Releases Wiki Activity

forgot to fix up the test trainer

Browse Source
This commit is contained in:
mrq 2024-04-21 14:58:04 -05:00
parent 071fb97777
commit b251669536
1 changed files with 15 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
vall_e/models/ar_nar.py
View File

@ -318,32 +318,36 @@ class AR_NAR(Base):
def example_usage(): def example_usage():
#cfg.trainer.backend = "local" #cfg.trainer.backend = "local"
from functools import partial cfg.hyperparameters.gradient_accumulation_steps = 1
from functools import partial
from einops import repeat from einops import repeat
from tqdm import tqdm
from ..emb.qnt import decode_to_file, unload_model from ..emb.qnt import decode_to_file, unload_model
from ..engines import Engine from ..engines import Engine
from tqdm import tqdm
from ..utils import wrapper as ml from ..utils import wrapper as ml
import numpy as np
import re import re
device = "cuda" device = "cuda"
x8 = partial(repeat, pattern="t -> t l", l=cfg.model.prom_levels) x8 = partial(repeat, pattern="t -> t l", l=cfg.model.prom_levels)
symmap = {'<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}
def tokenize(content, lang_marker="en"):
split = content.split(" ")
phones = [f"<s>"] + [ " " if not p else p for p in split ] + [f"</s>"]
return torch.tensor([*map(symmap.get, phones)])
qnt = torch.load(f'data/qnt{".dac" if cfg.inference.audio_backend == "dac" else ""}.pt')[0].t()[:, :cfg.model.prom_levels].to(device) def tokenize(content):
return torch.tensor( cfg.tokenizer.encode(content) )
print(qnt.shape) def _load_quants(path) -> Tensor:
if cfg.inference.audio_backend == "dac":
qnt = np.load(f'{path}.dac', allow_pickle=True)[()]
return torch.from_numpy(qnt["codes"].astype(int))[0][:, :].t().to(torch.int16)
return torch.load(f'{path}.pt')[0][:, :cfg.model.prom_levels].t().to(torch.int16)
qnt = _load_quants("./data/qnt")
cfg.hyperparameters.gradient_accumulation_steps = 1
text_list = [ 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("ˈaɪ wɪl nˌɑːt ˈæsk ɐ sˈɛkənd tˈaɪm").to(device),
] ]
proms_list = [ proms_list = [
qnt[:75*3, :].to(device), qnt[:75*3, :].to(device),