def synthesis(text, args): m = Model() m_post = ModelPostNet() m.load_state_dict(load_checkpoint(args.restore_step1, "transformer")) m_post.load_state_dict(load_checkpoint(args.restore_step2, "postnet")) text = np.asarray(text_to_sequence(text, [hp.cleaners])) text = t.LongTensor(text).unsqueeze(0) text = text.cuda() mel_input = t.zeros([1,1, 80]).cuda() pos_text = t.arange(1, text.size(1)+1).unsqueeze(0) pos_text = pos_text.cuda() m=m.cuda() m_post = m_post.cuda() m.train(False) m_post.train(False) pbar = tqdm(range(args.max_len)) with t.no_grad(): for i in pbar: pos_mel = t.arange(1,mel_input.size(1)+1).unsqueeze(0).cuda() mel_pred, postnet_pred, attn, stop_token, _, attn_dec = m.forward(text, mel_input, pos_text, pos_mel) mel_input = t.cat([t.zeros([1,1, 80]).cuda(),postnet_pred], dim=1) mag_pred = m_post.forward(postnet_pred) wav = spectrogram2wav(mag_pred.squeeze(0).cpu().numpy()) write(hp.sample_path + "/test.wav", hp.sr, wav)
def synthesis(text, num): m = Model() # m_post = ModelPostNet() m.load_state_dict(load_checkpoint(num, "transformer")) # m_post.load_state_dict(load_checkpoint(args.restore_step2, "postnet")) text = np.asarray(text_to_sequence(text, [hp.cleaners])) text = t.LongTensor(text).unsqueeze(0) text = text.cuda() mel_input = t.zeros([1, 1, 80]).cuda() pos_text = t.arange(1, text.size(1) + 1).unsqueeze(0) pos_text = pos_text.cuda() m = m.cuda() # m_post = m_post.cuda() m.train(False) # m_post.train(False) # pbar = tqdm(range(args.max_len)) with t.no_grad(): for _ in range(1000): pos_mel = t.arange(1, mel_input.size(1) + 1).unsqueeze(0).cuda() mel_pred, postnet_pred, attn, stop_token, _, attn_dec = m.forward( text, mel_input, pos_text, pos_mel) mel_input = t.cat([mel_input, postnet_pred[:, -1:, :]], dim=1) # mag_pred = m_post.forward(postnet_pred) # wav = spectrogram2wav(mag_pred.squeeze(0).cpu().numpy()) mel_postnet = postnet_pred[0].cpu().numpy().T plot_data([mel_postnet for _ in range(2)]) wav = audio.inv_mel_spectrogram(mel_postnet) wav = wav[0:audio.find_endpoint(wav)] audio.save_wav(wav, "result.wav")
def synthesis(text, args): m = Model() m.load_state_dict(load_checkpoint(args.restore_path)) print("[%s][%s] Synthesizing:" % (args.lang, args.spk), text) text = np.asarray([1] + list(text.encode('utf-8')) + [2]) text = t.LongTensor(text).unsqueeze(0) text = text mel_input = t.zeros([1, 1, 80]) pos_text = t.arange(1, text.size(1) + 1).unsqueeze(0) pos_text = pos_text lang_to_id = json.load(open(os.path.join(args.data_path, 'lang_id.json'))) spk_to_id = json.load(open(os.path.join(args.data_path, 'spk_id.json'))) lang_id = lang_to_id[args.lang] spk_id = spk_to_id[args.spk] lang_id = t.LongTensor([lang_id]) spk_id = t.LongTensor([spk_id]) m.train(False) pbar = tqdm(range(args.max_len)) with t.no_grad(): for i in pbar: pos_mel = t.arange(1, mel_input.size(1) + 1).unsqueeze(0) mel_pred, postnet_pred, attn, stop_token, _, attn_dec = \ m.forward(text, mel_input, pos_text, pos_mel, lang_id, spk_id) mel_input = t.cat([mel_input, mel_pred[:, -1:, :]], dim=1) if stop_token[:, -1].item() > 0: break mel = postnet_pred.squeeze(0).cpu().numpy() wav = mel2wav(mel) np.save(args.out_path + "_mel.npy", mel) write(args.out_path + ".wav", hp.sr, wav) plot_mel(args.out_path + "_mel.png", mel) plot_attn(attn, args.out_path + '_align.png')
def synthesis(text, args, num): m = Model() m_post = ModelPostNet() m.load_state_dict(load_checkpoint(args.restore_step1, "transformer")) m_post.load_state_dict(load_checkpoint(args.restore_step2, "postnet")) text = np.asarray(text_to_sequence(text, [hp.cleaners])) text = t.LongTensor(text).unsqueeze(0) text = text.cuda() mel_input = t.zeros([1, 1, 80]).cuda() pos_text = t.arange(1, text.size(1) + 1).unsqueeze(0) pos_text = pos_text.cuda() m = m.cuda() m_post = m_post.cuda() m.train(False) m_post.train(False) pbar = tqdm(range(args.max_len)) with t.no_grad(): for i in pbar: pos_mel = t.arange(1, mel_input.size(1) + 1).unsqueeze(0).cuda() mel_pred, postnet_pred, attn, stop_token, _, attn_dec = m.forward( text, mel_input, pos_text, pos_mel) # print('mel_pred==================',mel_pred.shape) # print('postnet_pred==================', postnet_pred.shape) mel_input = t.cat([mel_input, postnet_pred[:, -1:, :]], dim=1) #print(postnet_pred[:, -1:, :]) #print(t.argmax(attn[1][1][i]).item()) #print('mel_input==================', mel_input.shape) # #直接用真实mel测试postnet效果 #aa = t.from_numpy(np.load('D:\SSHdownload\\000101.pt.npy')).cuda().unsqueeze(0) # # print(aa.shape) mag_pred = m_post.forward(postnet_pred) #real_mag = t.from_numpy((np.load('D:\SSHdownload\\003009.mag.npy'))).cuda().unsqueeze(dim=0) #wav = spectrogram2wav(postnet_pred) #print('shappe============',attn[2][0].shape) # count = 0 # for j in range(4): # count += 1 # attn1 = attn[0][j].cpu() # plot_alignment(attn1, path='./training_loss/'+ str(args.restore_step1)+'_'+str(count)+'_'+'S'+str(num)+'.png', title='sentence'+str(num)) attn1 = attn[0][1].cpu() plot_alignment(attn1, path='./training_loss/' + str(args.restore_step1) + '_' + 'S' + str(num) + '.png', title='sentence' + str(num)) wav = spectrogram2wav(mag_pred.squeeze(0).cpu().detach().numpy()) write( hp.sample_path + '/' + str(args.restore_step1) + '-' + "test" + str(num) + ".wav", hp.sr, wav)
def synthesis(args): m = Model() m_post = ModelPostNet() m_stop = ModelStopToken() m.load_state_dict(load_checkpoint(args.restore_step1, "transformer")) m_stop.load_state_dict(load_checkpoint(args.restore_step3, "stop_token")) m_post.load_state_dict(load_checkpoint(args.restore_step2, "postnet")) m=m.cuda() m_post = m_post.cuda() m_stop = m_stop.cuda() m.train(False) m_post.train(False) m_stop.train(False) test_dataset = get_dataset(hp.test_data_csv) test_dataloader = DataLoader(test_dataset, batch_size=1, shuffle=False, collate_fn=collate_fn_transformer, drop_last=True, num_workers=1) ref_dataset = get_dataset(hp.test_data_csv) ref_dataloader = DataLoader(ref_dataset, batch_size=1, shuffle=True, collate_fn=collate_fn_transformer, drop_last=True, num_workers=1) writer = get_writer(hp.checkpoint_path, hp.log_directory) ref_dataloader_iter = iter(ref_dataloader) for i, data in enumerate(test_dataloader): character, mel, mel_input, pos_text, pos_mel, text_length, mel_length, fname = data ref_character, ref_mel, ref_mel_input, ref_pos_text, ref_pos_mel, ref_text_length, ref_mel_length, ref_fname = next(ref_dataloader_iter) stop_tokens = t.abs(pos_mel.ne(0).type(t.float) - 1) mel_input = t.zeros([1,1,80]).cuda() stop=[] character = character.cuda() mel = mel.cuda() mel_input = mel_input.cuda() pos_text = pos_text.cuda() pos_mel = pos_mel.cuda() ref_character = ref_character.cuda() ref_mel = ref_mel.cuda() ref_mel_input = ref_mel_input.cuda() ref_pos_text = ref_pos_text.cuda() ref_pos_mel = ref_pos_mel.cuda() with t.no_grad(): start=time.time() for i in range(args.max_len): pos_mel = t.arange(1,mel_input.size(1)+1).unsqueeze(0).cuda() mel_pred, postnet_pred, attn_probs, decoder_output, attns_enc, attns_dec, attns_style = m.forward(character, mel_input, pos_text, pos_mel, ref_mel, ref_pos_mel) stop_token = m_stop.forward(decoder_output) mel_input = t.cat([mel_input, postnet_pred[:,-1:,:]], dim=1) stop.append(t.sigmoid(stop_token).squeeze(-1)[0,-1]) if stop[-1] > 0.5: print("stop token at " + str(i) + " is :", stop[-1]) print("model inference time: ", time.time() - start) break if stop[-1] == 0: continue mag_pred = m_post.forward(postnet_pred) inf_time = time.time() - start print("inference time: ", inf_time) wav = spectrogram2wav(mag_pred.squeeze(0).cpu().numpy()) print("rtx : ", (len(wav)/hp.sr) / inf_time) wav_path = os.path.join(hp.sample_path, 'wav') if not os.path.exists(wav_path): os.makedirs(wav_path) write(os.path.join(wav_path, "text_{}_ref_{}_synth.wav".format(fname, ref_fname)), hp.sr, wav) print("written as text{}_ref_{}_synth.wav".format(fname, ref_fname)) attns_enc_new=[] attns_dec_new=[] attn_probs_new=[] attns_style_new=[] for i in range(len(attns_enc)): attns_enc_new.append(attns_enc[i].unsqueeze(0)) attns_dec_new.append(attns_dec[i].unsqueeze(0)) attn_probs_new.append(attn_probs[i].unsqueeze(0)) attns_style_new.append(attns_style[i].unsqueeze(0)) attns_enc = t.cat(attns_enc_new, 0) attns_dec = t.cat(attns_dec_new, 0) attn_probs = t.cat(attn_probs_new, 0) attns_style = t.cat(attns_style_new, 0) attns_enc = attns_enc.contiguous().view(attns_enc.size(0), 1, hp.n_heads, attns_enc.size(2), attns_enc.size(3)) attns_enc = attns_enc.permute(1,0,2,3,4) attns_dec = attns_dec.contiguous().view(attns_dec.size(0), 1, hp.n_heads, attns_dec.size(2), attns_dec.size(3)) attns_dec = attns_dec.permute(1,0,2,3,4) attn_probs = attn_probs.contiguous().view(attn_probs.size(0), 1, hp.n_heads, attn_probs.size(2), attn_probs.size(3)) attn_probs = attn_probs.permute(1,0,2,3,4) attns_style = attns_style.contiguous().view(attns_style.size(0), 1, hp.n_heads, attns_style.size(2), attns_style.size(3)) attns_style = attns_style.permute(1,0,2,3,4) save_dir = os.path.join(hp.sample_path, 'figure', "text_{}_ref_{}_synth.wav".format(fname, ref_fname)) if not os.path.exists(save_dir): os.makedirs(save_dir) writer.add_alignments(attns_enc.detach().cpu(), attns_dec.detach().cpu(), attn_probs.detach().cpu(), attns_style.detach().cpu(), mel_length, text_length, args.restore_step1, 'Validation', save_dir)