import torch
def get_synthesizer(pth_path, device=torch.device("cpu")):
from infer.lib.infer_pack.models import (
SynthesizerTrnMs256NSFsid,
SynthesizerTrnMs256NSFsid_nono,
SynthesizerTrnMs768NSFsid,
SynthesizerTrnMs768NSFsid_nono,
)
cpt = torch.load(pth_path, map_location=torch.device("cpu"))
# tgt_sr = cpt["config"][-1]
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
if_f0 = cpt.get("f0", 1)
version = cpt.get("version", "v1")
if version == "v1":
if if_f0 == 1:
net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=False)
else:
net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
elif version == "v2":
net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=False)
net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
del net_g.enc_q
# net_g.forward = net_g.infer
# ckpt = {}
# ckpt["config"] = cpt["config"]
# ckpt["f0"] = if_f0
# ckpt["version"] = version
# ckpt["info"] = cpt.get("info", "0epoch")
net_g.load_state_dict(cpt["weight"], strict=False)
net_g = net_g.float()
net_g.eval().to(device)
net_g.remove_weight_norm()
return net_g, cpt