checkpoint = torch.load('test/TTSglow_130000')
model.load_state_dict(checkpoint['model'].state_dict())
dataset = FastSpeechDataset()
testing_loader = DataLoader(dataset,
batch_size=1,
shuffle=False,
collate_fn=collate_fn,
drop_last=True,
num_workers=4)
model = model.eval()
for i, data_of_batch in enumerate(testing_loader):
src_seq = data_of_batch["texts"]
src_pos = data_of_batch["pos"]
src_seq = torch.from_numpy(src_seq).long().to(device)
src_pos = torch.from_numpy(src_pos).long().to(device)
mel = model.inference(src_seq, src_pos, sigma=1.0, alpha=1.0)
mel = mel.squeeze()
print(mel.size())
mel_path = os.path.join("results", "{}_synthesis.pt".format(i))
torch.save(mel, mel_path)
plot_data([mel.cpu().numpy().T], i)
if i > 10:
break
''' glow = get_waveglow()
synthesis_waveglow(mel, glow, i, alpha=1.0)
print("Synthesized by Waveglow.")'''
src_seq = data_of_batch["texts"]
src_pos = data_of_batch["pos"]
mel_tgt = data_of_batch["mels"]
alignment_target = data_of_batch["alignment"]
src_seq = torch.from_numpy(src_seq).long().to(device)
src_pos = torch.from_numpy(src_pos).long().to(device)
mel_tgt = torch.from_numpy(mel_tgt).float().to(device)
alignment_target = torch.from_numpy(alignment_target).float().to(
device)
mel_max_len = mel_tgt.size(1)
with torch.no_grad():
audio = model.inference(src_seq,
src_pos,
mel_max_len,
alignment_target,
sigma=1.0,
alpha=1.0)
audio = audio * MAX_WAV_VALUE
audio = audio.squeeze()
print(torch.mean(audio))
audio = audio.cpu().numpy()
audio_tgt = torch.cat(audio_tgt)
audio_tgt = audio_tgt * MAX_WAV_VALUE
print(torch.mean(audio_tgt))
#print (audio_tgt)
audio_tgt = audio_tgt.squeeze()
audio_tgt = audio_tgt.cpu().numpy()
audio = audio.astype('int16')
