AIÆ×Çú | »ùÓÚRWKVµÄ×î¼Ñ¿ªÔ´AI×÷ÇúÄ£ÐÍħ´îÍÆÀíʵ¼ù

# °²×°git, git-lfs. ÕâÀïÌṩÁ½ÖÖ·½Ê½
sudo apt install git git-lfs  # ʹÓÃapt°²×°
# conda install git git-lfs -c conda-forge  # ʹÓÃconda°²×°
git clone http://www.modelscope.cn/studios/AI-ModelScope/RWKV-4-Music.git

pip install rwkv tqdm mido midi2audio requests certifi -U
# °²×°fluidsynthÈí¼þ°ü, ÔÚmidi->wavµÄʱºòÐèÒª. ÎÒÃǸø³öÁËÁ½ÖÖ°²×°·½Ê½. 
sudo apt install fluidsynth  # ʹÓÃapt°²×°
# conda install fluidsynth -c conda-forge  # ʹÓÃconda°²×°

import os
import torch
from rwkv.model import RWKV
from rwkv.utils import PIPELINE
from tqdm import tqdm
from modelscope import snapshot_download

model_dir = snapshot_download('AI-ModelScope/rwkv-4-music', revision='v1.0.1')
ckpt_120M_fname = "RWKV-4-MIDI-120M-v1-20230714-ctx4096.pth"
ckpt_560M_fname = "RWKV-4-MIDI-560M-v1-20230717-ctx4096.pth"
ckpt_fpath = os.path.join(model_dir, ckpt_560M_fname)
# tokenizer-midi.jsonÎļþÏÂÔØ·¾¶: https://github.com/BlinkDL/ChatRWKV/blob/main/music/tokenizer-midi.json
tokenizer_fpath = "tokenizer-midi.json"
model = RWKV(model=ckpt_fpath, strategy='cuda fp32')
model.eval()
pipeline = PIPELINE(model, tokenizer_fpath)

# e.g. 
# input_text = 'v:5b:3 v:5b:2 t125 t125 t125 t106 pi:43:5 t24 pi:4a:7 t15 pi:4f:7 t17 pi:56:7 t18 pi:54:7 t125 t49 pi:51:7 t117 pi:4d:7 t125 t125 t111 pi:37:7 t14 pi:3e:6 t15 pi:43:6 t12 pi:4a:7 t17 pi:48:7 t125 t60 pi:45:7 t121 pi:41:7 t125 t117 s:46:5 s:52:5 f:46:5 f:52:5 t121 s:45:5 s:46:0 s:51:5 s:52:0 f:45:5 f:46:0 f:51:5 f:52:0 t121 s:41:5 s:45:0 s:4d:5 s:51:0 f:41:5 f:45:0 f:4d:5 f:51:0 t102 pi:37:0 pi:3e:0 pi:41:0 pi:43:0 pi:45:0 pi:48:0 pi:4a:0 pi:4d:0 pi:4f:0 pi:51:0 pi:54:0 pi:56:0 t19 s:3e:5 s:41:0 s:4a:5 s:4d:0 f:3e:5 f:41:0 f:4a:5 f:4d:0 t121 v:3a:5 t121 v:39:7 t15 v:3a:0 t106 v:35:8 t10 v:39:0 t111 v:30:8 v:35:0 t125 t117 v:32:8 t10 v:30:0 t125 t125 t103 v:5b:0 v:5b:0 t9 pi:4a:7'
input_text = ''
only_piano = False
max_length = 512
max_length = min(max_length, 4096)
temperature = 1
top_k = 8
top_p = 0.8
#
txt_fpath = 'midi.txt'
midi_fpath = 'midi_bin.midi'
wav_fpath = 'wav_bin.wav'

input_text = input_text.strip()
input_text = f'<pad> {input_text}'
output_text = '<start>'
#
occurrence = {}
state = None
for i in tqdm(range(max_length)):
    if i == 0:
        out, state = model.forward(pipeline.encode(input_text), state)
    else:
        out, state = model.forward([token], state)
    #
    for n in occurrence:
        out[n] -= (0 + occurrence[n] * 0.5)
    out[0] += (i - 2000) / 500  # try not to be too short or too long
    out[127] -= 1  # avoid "t125"
    #
    if only_piano:
        out[128:12416] -= 1e10
        out[13952:20096] -= 1e10
    # find the best sampling for your taste
    token = pipeline.sample_logits(
        out, temperature=temperature, top_k=top_k, top_p=top_p)
    if token == 0:
        break
    for n in occurrence:
        occurrence[n] *= 0.997  # decay repetition penalty
    if token >= 128 or token == 127:
        occurrence[token] = 1 + \
            (occurrence[token] if token in occurrence else 0)
    else:
        occurrence[token] = 0.3 + \
            (occurrence[token] if token in occurrence else 0)
    output_text += f" {pipeline.decode([token])}"
output_text += ' <end>'

with open(txt_fpath, "w") as f:
    f.write(output_text)

# midi_utilÎļþÏÂÔØ: https://github.com/briansemrau/MIDI-LLM-tokenizer/blob/main/midi_util.py
# vocab_configÎļþÏÂÔØ: https://github.com/briansemrau/MIDI-LLM-tokenizer/blob/main/vocab_config.json
import midi_util
from midi_util import VocabConfig
def txt_to_midi(text_fpath: str, output_fpath: str):
    vocab_config = "vocab_config.json"
    cfg = VocabConfig.from_json(vocab_config)
    with open(text_fpath, "r") as f:
        text = f.read()
    text = text.strip()
    mid = midi_util.convert_str_to_midi(cfg, text)
    mid.save(output_fpath)
txt_to_midi(txt_fpath, midi_fpath)

# ÕâÀïµÄ.sf2ÎļþÏÂÔØ·¾¶: https://github.com/vyshor/MusicAids/blob/master/default_sound_font.sf2
from midi2audio import FluidSynth
def midi_to_wav(midi_path: str, wav_path: str) -> None:
    # ÉèÖÃÒôÔ´£¬ÄãÐèÒªÏÂÔØÒ»¸ö.sf2Îļþ£¬ÕâÊÇÒ»¸öÒôÔ´Îļþ
    fs = FluidSynth('default_sound_font.sf2')
    fs.midi_to_audio(midi_path, wav_path)
midi_to_wav(midi_fpath, wav_fpath)