xingjianll/aria-accompaniment

Model

Model Description

LoRA adapter for accompaniment generation, fine-tuned on ~20,000 (treble-clef, full song) pairs, based on Aria-Medium-Base. try the model here: https://huggingface.co/spaces/xingjianll/symbolic-music-gen

Get Started

  tokenizer = AutoTokenizer.from_pretrained(
      "loubb/aria-medium-base",
      trust_remote_code=True,
      add_eos_token=True,
      add_dim_token=False,
  )

  midi_dict = MidiDict.from_midi("input_midi_path")
  tokens = tokenizer.tokenize(
      midi_dict, add_eos_token=True, add_dim_token=False
  )
  token_ids = tokenizer._tokenizer.encode(tokens)
  input_ids = torch.tensor([token_ids], device='cpu')

  model = AutoModelForCausalLM.from_pretrained("loubb/aria-medium-base", trust_remote_code=True)
  model = PeftModel.from_pretrained(model, "xingjianll/aria-accompaniment")
  continuation = model.generate(
        input_ids,
        max_length=1600,
        do_sample=True,
        temperature=1,
        top_p=0.95,
        use_cache=True
    )
  midi_dict_output = tokenizer.decode(continuation[0][input_ids.shape[1]:].tolist())
  midi_dict_output.to_midi().save("out_midi_path")