171 lines
6.1 KiB
Python
Executable File
171 lines
6.1 KiB
Python
Executable File
import math
|
||
import torch
|
||
import torch.nn.functional as F
|
||
import traceback
|
||
|
||
from typing import Literal, overload
|
||
from functools import partial
|
||
from einops import rearrange
|
||
|
||
from torch import Tensor, einsum, nn
|
||
from torch.distributions import Categorical
|
||
from torch.nn.utils.rnn import pad_sequence
|
||
from torch.utils.checkpoint import checkpoint
|
||
from torchmetrics.classification import BinaryAccuracy, MulticlassAccuracy, MulticlassPrecision
|
||
from torchvision.models import resnet18
|
||
|
||
from ..data import get_symmap
|
||
|
||
def _create_mask(l, device):
|
||
"""1 is valid region and 0 is invalid."""
|
||
seq = torch.arange(max(l), device=device).unsqueeze(0) # (1 t)
|
||
stop = torch.tensor(l, device=device).unsqueeze(1) # (b 1)
|
||
return (seq < stop).float() # (b t)
|
||
|
||
def list_to_tensor(x_list: list[Tensor], pattern="t b c -> b t c"):
|
||
"""
|
||
Args:
|
||
x_list: [(t d)]
|
||
Returns:
|
||
x: (? ? ?)
|
||
m: (? ? ?), same as x
|
||
"""
|
||
l = list(map(len, x_list))
|
||
x = rearrange(pad_sequence(x_list), pattern)
|
||
m = _create_mask(l, x_list[0].device)
|
||
m = m.t().unsqueeze(-1) # (t b 1)
|
||
m = rearrange(m, pattern)
|
||
m = m.to(x)
|
||
return x, m
|
||
|
||
class Model(nn.Module):
|
||
def __init__(
|
||
self,
|
||
n_tokens: int = 0, # number of token types
|
||
n_len: int = 6, # how long a sequence can be
|
||
d_model: int = 512,
|
||
):
|
||
super().__init__()
|
||
|
||
|
||
_symmap = get_symmap()
|
||
self.symmap = { f'{v}': k for k, v in _symmap.items() }
|
||
self.symmap['0'] = ""
|
||
|
||
if n_tokens == 0:
|
||
n_tokens = len(_symmap.keys())
|
||
|
||
self.n_tokens = n_tokens
|
||
self.n_len = n_len + 2 # start/stop tokens
|
||
self.d_model = d_model
|
||
|
||
self.resnet = resnet18(pretrained=False)
|
||
self.resnet.fc = nn.Linear( self.d_model, self.n_tokens * self.n_len )
|
||
|
||
self.criterion = nn.CTCLoss(zero_infinity=True)
|
||
|
||
def forward(
|
||
self,
|
||
|
||
image,
|
||
text = None,
|
||
|
||
sampling_temperature: float = 1.0,
|
||
):
|
||
x_list = torch.stack( image, dim=0 )
|
||
|
||
x = self.resnet( x_list )
|
||
y = x.view(x.size(0), self.n_len, self.n_tokens)
|
||
|
||
# pred = y.argmax(dim=2)
|
||
pred = Categorical(logits=y / sampling_temperature).sample()
|
||
|
||
answer = [ "".join([ self.symmap[f'{x.item()}'] for x in t ]) for t in pred ]
|
||
|
||
if text is not None:
|
||
y_list = rearrange(pad_sequence(text), "t b -> b t")
|
||
|
||
loss = 0
|
||
for i in range(self.n_len):
|
||
loss += F.cross_entropy( y[:, i], y_list[:, i] )
|
||
|
||
self.loss = dict(
|
||
nll=loss
|
||
)
|
||
|
||
return answer
|
||
|
||
def example_usage():
|
||
from ..config import cfg
|
||
cfg.trainer.backend = "local"
|
||
cfg.trainer.check_for_oom = False
|
||
|
||
from functools import partial
|
||
|
||
from einops import repeat
|
||
|
||
from ..emb.qnt import decode_to_file
|
||
from ..engines import Engine, Engines
|
||
from tqdm import tqdm, trange
|
||
|
||
from .ar import AR
|
||
from .nar import NAR
|
||
|
||
device = "cpu"
|
||
x8 = partial(repeat, pattern="t -> t l", l=2)
|
||
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)]).to()
|
||
|
||
kwargs = {
|
||
'n_tokens': 1024,
|
||
'd_model': 1024,
|
||
'n_heads': 16,
|
||
'n_layers': 12,
|
||
}
|
||
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
|
||
|
||
|
||
|
||
|
||
def sample( name, steps=400 ):
|
||
AR = None
|
||
NAR = None
|
||
|
||
engines.eval()
|
||
for name, engine in engines.items():
|
||
if name[:2] == "ar":
|
||
AR = engine
|
||
elif name[:3] == "nar":
|
||
NAR = engine
|
||
|
||
resps_list = AR(text_list, proms_list, max_steps=steps, sampling_temperature=1.0)
|
||
resps_list = [r.unsqueeze(-1) for r in resps_list]
|
||
codes = NAR( text_list, proms_list, resps_list=resps_list, sampling_temperature=0.2 )
|
||
|
||
decode_to_file(resps_list[0], f"./data/ar.{name}.wav", device=device)
|
||
decode_to_file(codes[0], f"./data/ar+nar.{name}.wav", device=device)
|
||
|
||
if train:
|
||
sample("init", 15)
|
||
|
||
engines.train()
|
||
t = trange(60)
|
||
for i in t:
|
||
stats = engines.step({"text_list": text_list, "proms_list": proms_list, "resps_list": resps_list}, device="cpu")
|
||
t.set_description(f"{stats}")
|
||
else:
|
||
for name, engine in engines.items():
|
||
engine.module.load_state_dict(torch.load(f"./data/{name}.pth"))
|
||
|
||
sample("final")
|
||
|
||
|
||
if __name__ == "__main__":
|
||
example_usage()
|