Update use_gpt_tts to be usable with unified_voice2

codes/models/gpt_voice/transformer_builders.py

@@ -42,6 +42,9 @@ class LearnedPositionEmbeddings(nn.Module):
         sl = x.shape[1]
         return self.emb(torch.arange(0, sl, device=x.device))
 
+    def get_fixed_embedding(self, ind, dev):
+        return self.emb(torch.tensor([ind], device=dev)).unsqueeze(0)
+
 
 def build_hf_gpt_transformer(layers, model_dim, heads, max_mel_seq_len, max_text_seq_len, checkpointing):
     """

codes/models/gpt_voice/unified_voice2.py

@@ -3,10 +3,11 @@ import functools
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from transformers import GPT2Model, GPT2Config
+from transformers import GPT2Model, GPT2Config, GPT2PreTrainedModel
+from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions
+from transformers.utils.model_parallel_utils import get_device_map, assert_device_map
 
 from models.arch_util import AttentionBlock
-from models.gpt_voice.gpt_asr_hf import GPT2InferenceModel
 from models.gpt_voice.gpt_asr_hf2 import ResBlock
 from models.gpt_voice.transformer_builders import build_hf_gpt_transformer
 from models.tacotron2.text import symbols
@@ -14,6 +15,160 @@ from trainer.networks import register_model
 from utils.util import opt_get
 
 
+class GPT2InferenceModel(GPT2PreTrainedModel):
+    def __init__(self, config, gpt, text_pos_emb, embeddings, norm, linear):
+        super().__init__(config)
+        self.transformer = gpt
+        self.text_pos_embedding = text_pos_emb
+        self.embeddings = embeddings
+        self.lm_head = nn.Sequential(norm, linear)
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+        self.cached_mel_emb = None
+
+    def parallelize(self, device_map=None):
+        self.device_map = (
+            get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.transformer.h))
+        self.transformer.parallelize(self.device_map)
+        self.lm_head = self.lm_head.to(self.transformer.first_device)
+        self.model_parallel = True
+
+    def deparallelize(self):
+        self.transformer.deparallelize()
+        self.transformer = self.transformer.to("cpu")
+        self.lm_head = self.lm_head.to("cpu")
+        self.model_parallel = False
+        torch.cuda.empty_cache()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def store_mel_emb(self, mel_emb):
+        self.cached_mel_emb = mel_emb
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
+
+        token_type_ids = kwargs.get("token_type_ids", None)
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
+
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
+        else:
+            position_ids = None
+        return {
+            "input_ids": input_ids,
+            "past_key_values": past,
+            "use_cache": kwargs.get("use_cache"),
+            "position_ids": position_ids,
+            "attention_mask": attention_mask,
+            "token_type_ids": token_type_ids,
+        }
+
+    def forward(
+        self,
+        input_ids=None,
+        past_key_values=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        assert self.cached_mel_emb is not None
+        assert inputs_embeds is None  # Not supported by this inference model.
+        assert labels is None  # Training not supported by this inference model.
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # Create embedding
+        mel_len = self.cached_mel_emb.shape[1]
+        if input_ids.shape[1] != 1:
+            text_inputs = input_ids[:, mel_len:]
+            text_emb = self.embeddings(text_inputs)
+            text_emb = text_emb + self.text_pos_embedding(text_emb)
+            if self.cached_mel_emb.shape[0] != text_emb.shape[0]:
+                mel_emb = self.cached_mel_emb.repeat_interleave(text_emb.shape[0]//self.cached_mel_emb.shape[0], 0)
+            else:
+                mel_emb = self.cached_mel_emb
+            emb = torch.cat([mel_emb, text_emb], dim=1)
+        else:
+            emb = self.embeddings(input_ids)
+            emb = emb + self.text_pos_embedding.get_fixed_embedding(attention_mask.shape[1]-mel_len, attention_mask.device)
+
+        transformer_outputs = self.transformer(
+            inputs_embeds=emb,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.transformer.first_device)
+            hidden_states = hidden_states.to(self.lm_head.weight.device)
+
+        lm_logits = self.lm_head(hidden_states)
+
+        if not return_dict:
+            return (lm_logits,) + transformer_outputs[1:]
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=None,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+            cross_attentions=transformer_outputs.cross_attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(past, beam_idx):
+        """
+        This function is used to re-order the :obj:`past_key_values` cache if
+        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
+        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
+        """
+        return tuple(
+            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
+            for layer_past in past
+        )
+
+
 class ConditioningEncoder(nn.Module):
     def __init__(self,
                  spec_dim,
@@ -275,9 +430,9 @@ class UnifiedVoice(nn.Module):
         return loss_mel.mean()
 
     def inference_speech(self, speech_conditioning_input, text_inputs, **hf_generate_kwargs):
+        seq_length = self.max_mel_tokens + self.max_text_tokens + 2
         if not hasattr(self, 'inference_model'):
             # TODO: Decouple gpt_config from this inference model.
-            seq_length = self.max_mel_tokens + self.max_text_tokens  + 5
             gpt_config = GPT2Config(vocab_size=self.max_mel_tokens,
                                     n_positions=seq_length,
                                     n_ctx=seq_length,
@@ -286,7 +441,8 @@ class UnifiedVoice(nn.Module):
                                     n_head=self.heads,
                                     gradient_checkpointing=False,
                                     use_cache=True)
-            self.inference_model = GPT2InferenceModel(gpt_config, self.gpt, self.mel_pos_embedding, self.final_norm, self.mel_head)
+            self.inference_model = GPT2InferenceModel(gpt_config, self.gpt, self.mel_pos_embedding, self.mel_embedding, self.final_norm, self.mel_head)
+            self.gpt.wte = self.mel_embedding
 
         text_inputs = F.pad(text_inputs, (0, 1), value=self.stop_text_token)
         text_inputs, text_targets = self.build_aligned_inputs_and_targets(text_inputs, self.start_text_token, self.stop_text_token)
@@ -301,11 +457,11 @@ class UnifiedVoice(nn.Module):
         emb = torch.cat([conds, text_emb], dim=1)
         self.inference_model.store_mel_emb(emb)
 
-        fake_inputs = torch.full((emb.shape[0], emb.shape[1]+1,), fill_value=1, dtype=torch.long, device=text_inputs.device)
+        fake_inputs = torch.full((emb.shape[0], conds.shape[1]+emb.shape[1],), fill_value=1, dtype=torch.long, device=text_inputs.device)
         fake_inputs[:,-1] = self.start_mel_token
 
         gen = self.inference_model.generate(fake_inputs, bos_token_id=self.start_mel_token, pad_token_id=self.stop_mel_token, eos_token_id=self.stop_mel_token,
-                                            max_length=self.seq_length, **hf_generate_kwargs)
+                                            max_length=seq_length, **hf_generate_kwargs)
         return gen[:, fake_inputs.shape[1]:]
 
 

codes/scripts/audio/gen/use_gpt_tts.py

@@ -80,13 +80,13 @@ def fix_autoregressive_output(codes, stop_token):
 
 if __name__ == '__main__':
     preselected_cond_voices = {
-        'trump': 'D:\\data\\audio\\sample_voices\\trump.wav',
+        'trump': ['D:\\data\\audio\\sample_voices\\trump.wav'],
-        'ryan_reynolds': 'D:\\data\\audio\\sample_voices\\ryan_reynolds.wav',
+        'ryan_reynolds': ['D:\\data\\audio\\sample_voices\\ryan_reynolds.wav'],
-        'ed_sheeran': 'D:\\data\\audio\\sample_voices\\ed_sheeran.wav',
+        'ed_sheeran': ['D:\\data\\audio\\sample_voices\\ed_sheeran.wav'],
-        'simmons': 'Y:\\clips\\books1\\754_Dan Simmons - The Rise Of Endymion 356 of 450\\00026.wav',
+        'simmons': ['Y:\\clips\\books1\\754_Dan Simmons - The Rise Of Endymion 356 of 450\\00026.wav'],
-        'news_girl': 'Y:\\clips\\podcasts-0\\8288_20210113-Is More Violence Coming_\\00022.wav',
+        'news_girl': ['Y:\\clips\\podcasts-0\\8288_20210113-Is More Violence Coming_\\00022.wav', 'Y:\\clips\\podcasts-0\\8288_20210113-Is More Violence Coming_\\00016.wav'],
-        'dan_carlin': 'Y:\\clips\\books1\\5_dchha06 Shield of the West\\00476.wav',
+        'dan_carlin': ['Y:\\clips\\books1\\5_dchha06 Shield of the West\\00476.wav'],
-        'libri_test': 'Y:\\libritts\\test-clean\\672\\122797\\672_122797_000057_000002.wav'
+        'libri_test': ['Y:\\libritts\\test-clean\\672\\122797\\672_122797_000057_000002.wav']
     }
 
     parser = argparse.ArgumentParser()
@@ -94,17 +94,16 @@ if __name__ == '__main__':
     parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator')
     parser.add_argument('-diffusion_model_path', type=str, help='Diffusion model checkpoint to load.', default='X:\\dlas\\experiments\\train_diffusion_vocoder_with_cond_new_dvae_full\\models\\6100_generator_ema.pth')
     parser.add_argument('-dvae_model_name', type=str, help='Name of the DVAE model in opt.', default='dvae')
-    parser.add_argument('-opt_gpt_tts', type=str, help='Path to options YAML file used to train the GPT-TTS model', default='X:\\dlas\\experiments\\train_gpt_tts_unified\\train_gpt_tts_unified.yml')
+    parser.add_argument('-opt_gpt_tts', type=str, help='Path to options YAML file used to train the GPT-TTS model', default='X:\\dlas\\experiments\\train_gpt_tts_unified.yml')
     parser.add_argument('-gpt_tts_model_name', type=str, help='Name of the GPT TTS model in opt.', default='gpt')
-    parser.add_argument('-gpt_tts_model_path', type=str, help='GPT TTS model checkpoint to load.', default='X:\\dlas\\experiments\\train_gpt_tts_unified\\models\\60000_gpt_ema.pth')
+    parser.add_argument('-gpt_tts_model_path', type=str, help='GPT TTS model checkpoint to load.', default='X:\\dlas\\experiments\\train_gpt_tts_unified_large\\models\\40000_gpt_ema.pth')
     parser.add_argument('-opt_clip', type=str, help='Path to options YAML file used to train the CLIP model', default='X:\\dlas\\experiments\\train_clip_text_to_voice.yml')
     parser.add_argument('-clip_model_name', type=str, help='Name of the CLIP model in opt.', default='clip')
     parser.add_argument('-clip_model_path', type=str, help='CLIP model checkpoint to load.', default='X:\\dlas\\experiments\\train_clip_text_to_voice_masking_bigger_batch\\models\\23500_clip_ema.pth')
     parser.add_argument('-text', type=str, help='Text to speak.', default="I am a language model that has learned to speak.")
-    parser.add_argument('-cond_path', type=str, help='Path to condioning sample.', default='')
     parser.add_argument('-cond_preset', type=str, help='Use a preset conditioning voice (defined above). Overrides cond_path.', default='libri_test')
     parser.add_argument('-num_samples', type=int, help='How many total outputs the autoregressive transformer should produce.', default=128)
-    parser.add_argument('-num_batches', type=int, help='How many batches those samples should be produced over.', default=2)
+    parser.add_argument('-num_batches', type=int, help='How many batches those samples should be produced over.', default=8)
     parser.add_argument('-num_outputs', type=int, help='Number of outputs to produce.', default=2)
     parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='../results/use_gpt_tts')
     args = parser.parse_args()
@@ -115,7 +114,7 @@ if __name__ == '__main__':
     with open(args.opt_gpt_tts, mode='r') as f:
         gpt_opt = yaml.load(f, Loader=Loader)
     gpt_opt['networks'][args.gpt_tts_model_name]['kwargs']['checkpointing'] = False  # Required for beam search
-    gpt = load_model_from_config(preloaded_options=gpt_opt, model_name=args.gpt_tts_model_name, also_load_savepoint=False, load_path=args.gpt_tts_model_path, strict_load=False).eval()
+    gpt = load_model_from_config(preloaded_options=gpt_opt, model_name=args.gpt_tts_model_name, also_load_savepoint=False, load_path=args.gpt_tts_model_path).eval()
     stop_mel_token = gpt.stop_mel_token
 
     print("Loading data..")
@@ -123,8 +122,12 @@ if __name__ == '__main__':
     text = torch.IntTensor(tokenizer.encode(args.text)).unsqueeze(0).cuda()
     text = F.pad(text, (0,1))  # This may not be necessary.
 
-    cond_path = args.cond_path if args.cond_preset is None else preselected_cond_voices[args.cond_preset]
+    cond_paths = preselected_cond_voices[args.cond_preset]
-    conds, cond_wav = load_conditioning(cond_path, cond_length=88000)
+    conds = []
+    for cond_path in cond_paths:
+        c, cond_wav = load_conditioning(cond_path, cond_length=132300)
+        conds.append(c)
+    conds = torch.stack(conds, dim=1)  # And just use the last cond_wav for the diffusion model.
 
     with torch.no_grad():
         print("Performing GPT inference..")