| from pathlib import Path |
| import copy |
|
|
| import torch |
| import hydra |
| from omegaconf import OmegaConf |
| from transformers import PreTrainedModel, PretrainedConfig |
|
|
| class SemanticVocoderConfig(PretrainedConfig): |
| """Configuration class for SemanticVocoder model.""" |
| model_type = "semantic_vocoder" |
| |
| def __init__(self, |
| model_config=None, |
| **kwargs): |
| super().__init__(**kwargs) |
| self.model_config = model_config |
|
|
| class SemanticVocoder(PreTrainedModel): |
| """HuggingFace compatible SemanticVocoder model.""" |
| config_class = SemanticVocoderConfig |
| |
| def __init__(self, config): |
| super().__init__(config) |
|
|
| self.model = hydra.utils.instantiate(config.model_config) |
| |
| def forward(self, |
| content, |
| num_steps=100, |
| guidance_scale=3.5, |
| guidance_rescale=0.5, |
| vocoder_steps=200, |
| latent_shape=[768, 250], |
| **kwargs): |
| """Forward pass through the model.""" |
|
|
| waveform = self.model.inference( |
| content=[content], |
| condition=None, |
| task=["text_to_audio"], |
| num_steps=num_steps, |
| guidance_scale=guidance_scale, |
| guidance_rescale=guidance_rescale, |
| vocoder_steps=vocoder_steps, |
| latent_shape=latent_shape, |
| **kwargs, |
| ) |
| return waveform[0][0].cpu().numpy() |
|
|
