"WARNING: Text2Mel and SSRN have different saved configs. Will use Text2Mel config!"
)
Config.set_config(config_t2m)
# Load networks
print("Loading Text2Mel...")
text2mel = Text2Mel().to(device)
text2mel.eval()
text2mel_step = state_t2m["global_step"]
text2mel.load_state_dict(state_t2m["model"])
print("Loading SSRN...")
ssrn = SSRN().to(device)
ssrn.eval()
ssrn_step = state_ssrn["global_step"]
ssrn.load_state_dict(state_ssrn["model"])
while True:
text = input("> ")
text = spell_out_numbers(text, args.language)
text = normalize(text)
text = text + Config.vocab_end_of_text
text = vocab_lookup(text)
L = torch.tensor(text, device=device, requires_grad=False).unsqueeze(0)
S = torch.zeros(1,
Config.max_T,
Config.F,
requires_grad=False,
device=device)
previous_position = torch.zeros(1,
global_step = 0
# Learning rate decay. Noam scheme
warmup_steps = 4000.0
def decay(_):
step = global_step + 1
return warmup_steps**0.5 * min(step * warmup_steps**-1.5, step**-0.5)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=decay)
if args.restore_path is not None:
print("Restoring from checkpoint: {}".format(args.restore_path))
state = torch.load(args.restore_path, map_location=device)
global_step = state["global_step"]
net.load_state_dict(state["model"])
optimizer.load_state_dict(state["optimizer"])
scheduler.load_state_dict(state["scheduler"])
l1_criterion.load_state_dict(state["l1_criterion"])
bd_criterion.load_state_dict(state["bd_criterion"])
print("Loading dataset...")
dataset = TTSDataset(args.text_path,
args.mel_path,
args.lin_path,
data_in_memory=True)
batch_sampler = BucketBatchSampler(
inputs=[d["text"] for d in dataset.data],
batch_size=args.batch_size,
bucket_boundaries=[i for i in range(1, Config.max_N - 1, 20)])
data_loader = FastDataLoader(dataset,