def build_default_vocoder(self):
from fairseq.models.text_to_speech.vocoder import get_vocoder
vocoder = get_vocoder(self.args, self.data_cfg)
if torch.cuda.is_available() and not self.args.cpu:
vocoder = vocoder.cuda()
else:
vocoder = vocoder.cpu()
return vocoder
def build_generator(
self,
models,
args,
seq_gen_cls=None,
extra_gen_cls_kwargs=None,
):
if not self.args.target_is_code or self.args.eval_inference:
from fairseq.models.text_to_speech.vocoder import get_vocoder
self.vocoder = get_vocoder(self.args, self.data_cfg)
self.vocoder = (
self.vocoder.cuda()
if torch.cuda.is_available() and not self.args.cpu
else self.vocoder.cpu()
)
if self.args.target_is_code:
if self.args.n_frames_per_step == 1:
seq_generator = super().build_generator(
models,
args,
seq_gen_cls=None,
extra_gen_cls_kwargs=extra_gen_cls_kwargs,
)
else:
assert (
getattr(args, "beam", 1) == 1 and getattr(args, "nbest", 1) == 1
), "only support viterbi search for stacked units"
seq_generator = StackUnitSequenceGenerator(
self.tgt_dict,
self.args.target_code_size,
)
else:
if getattr(args, "teacher_forcing", False):
from fairseq.speech_generator import (
TeacherForcingAutoRegressiveSpeechGenerator,
)
generator = TeacherForcingAutoRegressiveSpeechGenerator
logger.info("Teacher forcing mode for generation")
else:
from fairseq.speech_generator import AutoRegressiveSpeechGenerator
generator = AutoRegressiveSpeechGenerator
seq_generator = generator(
models[0],
self.vocoder,
self.data_cfg,
max_iter=self.args.max_target_positions,
eos_prob_threshold=self.args.eos_prob_threshold,
)
return seq_generator