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_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, args): m = Model() m_post = ModelPostNet() m.load_state_dict(load_checkpoint(args.step1, "transformer")) m_post.load_state_dict(load_checkpoint(args.step2, "postnet")) text = np.asarray(text_to_sequence(text, [hp.cleaners])) text = torch.LongTensor(text).unsqueeze(0) text = text.cuda() mel_input = np.load('3_0.pt.npy') pos_text = torch.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) with torch.no_grad(): mag_pred = m_post.forward( torch.from_numpy(mel_input).unsqueeze(0).cuda()) wav = spectrogram2wav(mag_pred.squeeze(0).cpu().numpy()) write(hp.sample_path + "/test.wav", hp.sr, wav)
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)
def synthesis(args): m = Model() m.load_state_dict(load_checkpoint(args.restore_step1, "transformer")) m = m.cuda() m.train(False) vocoder = SmartVocoder(Hyperparameters(parse_args())) vocoder.load_state_dict( t.load('./mel2audio/merged_STFT_checkpoint.pth')["state_dict"]) vocoder = vocoder.cuda() vocoder.eval() with open('./hifi_gan/config.json') as f: data = f.read() json_config = json.loads(data) h = AttrDict(json_config) hifi_gan = Generator(h).cuda() state_dict_g = t.load('./hifi_gan/g_00334000', map_location='cuda') hifi_gan.load_state_dict(state_dict_g['generator']) hifi_gan.eval() hifi_gan.remove_weight_norm() 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_shuf) ref_dataloader = DataLoader(ref_dataset, batch_size=1, shuffle=False, collate_fn=collate_fn_transformer, drop_last=True, num_workers=1) writer = get_writer(hp.checkpoint_path, hp.log_directory) mel_basis = t.from_numpy( librosa.filters.mel(hp.sr, hp.n_fft, hp.n_mels, 50, 11000)).unsqueeze(0) # (n_mels, 1+n_fft//2) ref_dataloader_iter = iter(ref_dataloader) _, ref_mel, _, _, _, ref_pos_mel, _, _, ref_fname = next( ref_dataloader_iter) for i, data in enumerate(test_dataloader): character, _, _, _, pos_text, _, text_length, _, fname = data mel_input = t.zeros([1, 1, 80]).cuda() character = character.cuda() ref_mel = ref_mel.cuda() mel_input = mel_input.cuda() pos_text = pos_text.cuda() with t.no_grad(): start = time.time() memory, c_mask, attns_enc, duration_mask = m.encoder(character, pos=pos_text) style, coarse_emb = m.ref_encoder(ref_mel) memory = t.cat((memory, coarse_emb.expand(-1, memory.size(1), -1)), -1) memory = m.memory_coarse_layer(memory) duration_predictor_output = m.duration_predictor( memory, duration_mask) duration = t.ceil(duration_predictor_output) duration = duration * duration_mask # max_length = t.sum(duration).type(t.LongTensor) # print("length : ", max_length) monotonic_interpolation, pos_mel_, weights = m.length_regulator( memory, duration, duration_mask) kv_mask = t.zeros([1, mel_input.size(1), character.size(1)]).cuda() # B, t', N kv_mask[:, :, :3] = 1 kv_mask = kv_mask.eq(0) stop_flag = False ctr = 0 for j in range(1200): pos_mel = t.arange(1, mel_input.size(1) + 1).unsqueeze(0).cuda() mel_pred, postnet_pred, attn_probs, decoder_output, attns_dec, attns_style = m.decoder( memory, style, mel_input, c_mask, pos=pos_mel, ref_pos=ref_pos_mel, mono_inter=monotonic_interpolation[:, :mel_input.shape[1]], kv_mask=kv_mask) mel_input = t.cat([mel_input, postnet_pred[:, -1:, :]], dim=1) # print("j", j, "mel_input", mel_input.shape) if stop_flag and ctr == 10: break elif stop_flag: ctr += 1 kv_mask, stop_flag = update_kv_mask( kv_mask, attn_probs) # B, t', N --> B, t'+1, N postnet_pred = t.cat((postnet_pred, t.zeros(postnet_pred.size(0), 5, postnet_pred.size(-1)).cuda()), 1) gen_length = mel_input.size(1) # print("gen_length", gen_length) post_linear = m.postnet(postnet_pred) post_linear = resample(post_linear, seq_len=mel_input.size(1), scale=args.rhythm_scale) postnet_pred = resample(mel_input, seq_len=mel_input.size(1), scale=args.rhythm_scale) inf_time = time.time() - start print("inference time: ", inf_time) # print("speech_rate: ", len(postnet_pred[0])/len(character[0])) postnet_pred_v = postnet_pred.transpose(2, 1) postnet_pred_v = (postnet_pred_v * 100 + 20 - 100) / 20 B, C, T = postnet_pred_v.shape z = t.randn(1, 1, T * hp.hop_length).cuda() z = z * 0.6 # Temp t.cuda.synchronize() timestemp = time.time() with t.no_grad(): y_gen = vocoder.reverse(z, postnet_pred_v).squeeze() t.cuda.synchronize() print('{} seconds'.format(time.time() - timestemp)) wav = y_gen.to(t.device("cpu")).data.numpy() wav = np.pad( wav, [0, 4800], mode='constant', constant_values=0) #pad 0 for 0.21 sec silence at the end post_linear_v = post_linear.transpose(1, 2) post_linear_v = 10**((post_linear_v * 100 + 20 - 100) / 20) mel_basis = mel_basis.repeat(post_linear_v.shape[0], 1, 1) post_linear_mel_v = t.log10(t.bmm(mel_basis.cuda(), post_linear_v)) B, C, T = post_linear_mel_v.shape z = t.randn(1, 1, T * hp.hop_length).cuda() z = z * 0.6 # Temp t.cuda.synchronize() timestemp = time.time() with t.no_grad(): y_gen_linear = vocoder.reverse(z, post_linear_mel_v).squeeze() t.cuda.synchronize() wav_linear = y_gen_linear.to(t.device("cpu")).data.numpy() wav_linear = np.pad( wav_linear, [0, 4800], mode='constant', constant_values=0) #pad 0 for 0.21 sec silence at the end wav_hifi = hifi_gan(post_linear_mel_v).squeeze().clamp( -1, 1).detach().cpu().numpy() wav_hifi = np.pad( wav_hifi, [0, 4800], mode='constant', constant_values=0) #pad 0 for 0.21 sec silence at the end mel_path = os.path.join(hp.sample_path + '_' + str(args.rhythm_scale), 'mel') if not os.path.exists(mel_path): os.makedirs(mel_path) np.save( os.path.join( mel_path, 'text_{}_ref_{}_synth_{}.mel'.format(i, ref_fname, str(args.rhythm_scale))), postnet_pred.cpu()) linear_path = os.path.join( hp.sample_path + '_' + str(args.rhythm_scale), 'linear') if not os.path.exists(linear_path): os.makedirs(linear_path) np.save( os.path.join( linear_path, 'text_{}_ref_{}_synth_{}.linear'.format( i, ref_fname, str(args.rhythm_scale))), post_linear.cpu()) wav_path = os.path.join(hp.sample_path + '_' + str(args.rhythm_scale), 'wav') if not os.path.exists(wav_path): os.makedirs(wav_path) write( os.path.join( wav_path, "text_{}_ref_{}_synth_{}.wav".format(i, ref_fname, str(args.rhythm_scale))), hp.sr, wav) print("rtx : ", (len(wav) / hp.sr) / inf_time) wav_linear_path = os.path.join( hp.sample_path + '_' + str(args.rhythm_scale), 'wav_linear') if not os.path.exists(wav_linear_path): os.makedirs(wav_linear_path) write( os.path.join( wav_linear_path, "text_{}_ref_{}_synth_{}.wav".format(i, ref_fname, str(args.rhythm_scale))), hp.sr, wav_linear) wav_hifi_path = os.path.join( hp.sample_path + '_' + str(args.rhythm_scale), 'wav_hifi') if not os.path.exists(wav_hifi_path): os.makedirs(wav_hifi_path) write( os.path.join( wav_hifi_path, "text_{}_ref_{}_synth_{}.wav".format(i, ref_fname, str(args.rhythm_scale))), hp.sr, wav_hifi) show_weights = weights.contiguous().view(weights.size(0), 1, 1, weights.size(1), weights.size(2)) 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 + '_' + str(args.rhythm_scale), 'figure', "text_{}_ref_{}_synth_{}.wav".format(fname, ref_fname, str(args.rhythm_scale))) 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(), show_weights.detach().cpu(), [t.tensor(gen_length).type(t.LongTensor)], text_length, args.restore_step1, 'Inference', save_dir)
parameters['GPUs'] = (parameters['GPUs'], ) testset = MyDataset( filelist='../dataset/wp1_real.txt', input_transform=transforms.Compose([ Resize((300, 300)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) ) model = Model(parameters['n_classes']) model.load_state_dict(torch.load('wp1-cold.pth')) if parameters['GPUs']: model = model.cuda(parameters['GPUs'][0]) if len(parameters['GPUs']) > 1: model = nn.DataParallel(model, device_ids=parameters['GPUs']) model.eval() all_features, all_outputs, all_preds, all_labels = predict(model, testset, **parameters) recall = np.sum(all_preds == all_labels) / float(len(testset)) ap = AP(all_outputs, all_labels) mean_ap = meanAP(all_outputs, all_labels) print('Mean Recall: ', recall) print('AP of each class: ', ap) print('mean AP: ', mean_ap)