diff --git a/api.py b/api.py index 7c33484..e57ed03 100644 --- a/api.py +++ b/api.py @@ -133,7 +133,7 @@ class TextToSpeech: self.tokenizer = VoiceBpeTokenizer() download_models() - self.autoregressive = UnifiedVoice(max_mel_tokens=300, max_text_tokens=200, max_conditioning_inputs=2, layers=30, + self.autoregressive = UnifiedVoice(max_mel_tokens=604, max_text_tokens=402, max_conditioning_inputs=2, layers=30, model_dim=1024, heads=16, number_text_tokens=256, start_text_token=255, checkpointing=False, train_solo_embeddings=False, @@ -151,14 +151,18 @@ class TextToSpeech: layer_drop=0, unconditioned_percentage=0).cpu().eval() self.diffusion.load_state_dict(torch.load('.models/diffusion.pth')) + self.diffusion_next = DiffusionTts(model_channels=1024, num_layers=10, in_channels=100, out_channels=200, + in_latent_channels=1024, in_tokens=8193, dropout=0, use_fp16=False, num_heads=16, + layer_drop=0, unconditioned_percentage=0).cpu().eval() + self.diffusion_next.load_state_dict(torch.load('.models/diffusion_next.pth')) + self.vocoder = UnivNetGenerator().cpu() self.vocoder.load_state_dict(torch.load('.models/vocoder.pth')['model_g']) self.vocoder.eval(inference=True) def tts(self, text, voice_samples, k=1, # autoregressive generation parameters follow - num_autoregressive_samples=512, temperature=.5, length_penalty=2, repetition_penalty=2.0, top_p=.5, - typical_sampling=False, typical_mass=.9, + num_autoregressive_samples=512, temperature=.5, length_penalty=1, repetition_penalty=2.0, top_p=.5, # diffusion generation parameters follow diffusion_iterations=100, cond_free=True, cond_free_k=2, diffusion_temperature=.7,): text = torch.IntTensor(self.tokenizer.encode(text)).unsqueeze(0).cuda() @@ -185,10 +189,8 @@ class TextToSpeech: temperature=temperature, num_return_sequences=self.autoregressive_batch_size, length_penalty=length_penalty, - repetition_penalty=repetition_penalty, - typical_sampling=typical_sampling, - typical_mass=typical_mass) - padding_needed = 250 - codes.shape[1] + repetition_penalty=repetition_penalty) + padding_needed = self.autoregressive.max_mel_tokens - codes.shape[1] codes = F.pad(codes, (0, padding_needed), value=stop_mel_token) samples.append(codes) self.autoregressive = self.autoregressive.cpu() diff --git a/api_new_autoregressive.py b/api_new_autoregressive.py index 1ba90e4..cd5cd89 100644 --- a/api_new_autoregressive.py +++ b/api_new_autoregressive.py @@ -135,7 +135,7 @@ class TextToSpeech: download_models() self.autoregressive = AutoregressiveCodegen(1024, 16).cpu().eval() - self.autoregressive.load_state_dict(torch.load('X:\\dlas\\experiments\\train_autoregressive_codegen\\models\\17000_codegen_ema.pth')) + self.autoregressive.load_state_dict(torch.load('X:\\dlas\\experiments\\train_autoregressive_codegen\\models\\20750_codegen_ema.pth')) self.clip = VoiceCLIP(dim_text=512, dim_speech=512, dim_latent=512, num_text_tokens=256, text_enc_depth=12, text_seq_len=350, text_heads=8, diff --git a/eval_multiple.py b/eval_multiple.py index a3bf49f..99e1eae 100644 --- a/eval_multiple.py +++ b/eval_multiple.py @@ -7,7 +7,7 @@ from utils.audio import load_audio if __name__ == '__main__': fname = 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv' - outpath = 'D:\\tmp\\tortoise-tts-eval\\attempt_best' + outpath = 'D:\\tmp\\tortoise-tts-eval\\compare_vocoders' outpath_real = 'D:\\tmp\\tortoise-tts-eval\\real' os.makedirs(outpath, exist_ok=True) @@ -24,12 +24,18 @@ if __name__ == '__main__': path = os.path.join(os.path.dirname(fname), line[1]) cond_audio = load_audio(path, 22050) torchaudio.save(os.path.join(outpath_real, os.path.basename(line[1])), cond_audio, 22050) - sample = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=512, k=1, + sample, sample2 = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=512, k=1, repetition_penalty=2.0, length_penalty=2, temperature=.5, top_p=.5, - diffusion_temperature=.7, cond_free_k=2, diffusion_iterations=400) + diffusion_temperature=.7, cond_free_k=2, diffusion_iterations=200) + down = torchaudio.functional.resample(sample, 24000, 22050) - fout_path = os.path.join(outpath, os.path.basename(line[1])) + fout_path = os.path.join(outpath, 'old', os.path.basename(line[1])) torchaudio.save(fout_path, down.squeeze(0), 22050) + + down = torchaudio.functional.resample(sample2, 24000, 22050) + fout_path = os.path.join(outpath, 'new', os.path.basename(line[1])) + torchaudio.save(fout_path, down.squeeze(0), 22050) + recorder.write(f'{transcript}\t{fout_path}\n') recorder.flush() recorder.close() \ No newline at end of file diff --git a/models/new_autoregressive.py b/models/new_autoregressive.py index c372f62..aba8c11 100644 --- a/models/new_autoregressive.py +++ b/models/new_autoregressive.py @@ -168,6 +168,8 @@ class AutoregressiveCodegen(nn.Module): self.START_TOKEN=8192 self.STOP_TOKEN=8193 + self.START_TEXT_TOKEN = 255 + self.STOP_TEXT_TOKEN = 0 self.max_text_token_id = num_text_tokens self.max_mel_token_id = num_mel_tokens self.mel_embedding = ConditioningEncoder(80, model_dim, do_checkpointing=False) @@ -231,6 +233,9 @@ class AutoregressiveCodegen(nn.Module): for i in range(conditioning_signal.shape[1]): cond_embs.append(self.mel_embedding(conditioning_signal[:, i])) cond_emb = torch.stack(cond_embs, dim=1).mean(dim=1, keepdim=True) + # Since all positional embeddings are relative, it is (probably) important to "fix" the text with some permanent embeddings. + text_codes = F.pad(text_codes, (1,0), value=self.START_TEXT_TOKEN) + text_codes = F.pad(text_codes, (0,1), value=self.STOP_TEXT_TOKEN) _, enc_text = self.encoder(text_codes, return_hiddens=True) # Interleave cond_emb into the first few contexts. full_context = enc_text @@ -255,6 +260,8 @@ class AutoregressiveCodegen(nn.Module): for i in range(conditioning_signal.shape[1]): cond_embs.append(self.mel_embedding(conditioning_signal[:, i])) cond_emb = torch.stack(cond_embs, dim=1).mean(dim=1, keepdim=True) + text_codes = F.pad(text_codes, (1,0), value=self.START_TEXT_TOKEN) + text_codes = F.pad(text_codes, (0,1), value=self.STOP_TEXT_TOKEN) _, enc_text = self.encoder(text_codes, return_hiddens=True) # Interleave cond_emb into the first few contexts. full_context = enc_text diff --git a/models/text_voice_clip.py b/models/text_voice_clip.py index b4b51a7..674e62b 100644 --- a/models/text_voice_clip.py +++ b/models/text_voice_clip.py @@ -55,7 +55,6 @@ class VoiceCLIP(nn.Module): needs_permute=False, exit_permute=False, max_seq_len=-1, - use_pos_emb=False, attn_layers=Encoder( dim=dim_text, depth=text_enc_depth, @@ -71,7 +70,6 @@ class VoiceCLIP(nn.Module): needs_permute=False, exit_permute=False, max_seq_len=-1, - use_pos_emb=False, attn_layers=Encoder( dim=dim_speech, depth=speech_enc_depth, diff --git a/models/xtransformers.py b/models/xtransformers.py index 632349b..2e32c09 100644 --- a/models/xtransformers.py +++ b/models/xtransformers.py @@ -1186,7 +1186,9 @@ class TransformerWrapper(nn.Module): if use_cache: res.append(intermediates.past_key_values) - return res + if len(res) > 1: + return tuple(res) + return res[0] class ContinuousTransformerWrapper(nn.Module): @@ -1247,7 +1249,9 @@ class ContinuousTransformerWrapper(nn.Module): if use_cache: res.append(intermediates.past_key_values) - return tuple(res) + if len(res) > 1: + return tuple(res) + return res[0] class XTransformer(nn.Module):