Method untokenize

model.py:277–289

(self, token_data)

    @T.no_grad()
    def untokenize(self, token_data):
        if exists(self.audio_latent_cache):
            token_data = T.cat([self.audio_latent_cache, token_data], dim=1)
            self.audio_latent_cache = token_data[:, -(6*8):]
        elif self.use_audio_cache:
            self.audio_latent_cache = token_data[:, -(6*8):]
        
        if token_data.shape[-1] == 2*self.c.latent_size:
            dec_ch1 = self.audio_tokenizer.data_from_latent(token_data[:, :self.c.latent_size])
            dec_ch2 = self.audio_tokenizer.data_from_latent(token_data[:, self.c.latent_size:])
            return T.cat([dec_ch1, dec_ch2], dim=1)[..., -(token_data.shape[1]*2000):]
        else:
            return self.audio_tokenizer.data_from_latent(token_data)[..., -(token_data.shape[1]*2000):]

    def init_cache(self, bsize, device, dtype, length:int=None):
        cache_shape = [self.c.stack_config.layers, length or self.c.stack_config.seq_len, 2, self.kv_heads, self.head_dim]