musiclib.py

# apache 2.0 license, modified by mrfakename, from https://github.com/BlinkDL/ChatRWKV/tree/main/music
import torch
import os, sys
import numpy as np
from cached_path import cached_path
np.set_printoptions(precision=4, suppress=True, linewidth=200)

os.environ['RWKV_JIT_ON'] = '1' #### set these before import RWKV
os.environ["RWKV_CUDA_ON"] = '0'
os.environ["RWKV_RESCALE_LAYER"] = '999' # must set this for RWKV-music models and "pip install rwkv --upgrade" to v0.8.12+

from rwkv.model import RWKV
from rwkv.utils import PIPELINE

MODEL_FILE = str(cached_path('hf://BlinkDL/rwkv-4-music/RWKV-4-MIDI-120M-v1-20230714-ctx4096.pth'))

ABC_MODE = ('-ABC-' in MODEL_FILE)
MIDI_MODE = ('-MIDI-' in MODEL_FILE)
device = 'cpu'
if torch.cuda.is_available(): device = 'cuda'
model = RWKV(model=MODEL_FILE, strategy=f'{device} fp32')
pipeline = PIPELINE(model, "tokenizer-midi.json")

tokenizer = pipeline
EOS_ID = 0
TOKEN_SEP = ' '

def musicgen(ccc='<pad>', piano_only=False, length=4096):
    # ccc = '<pad>'
    ccc_output = '<start>'
    # ccc = "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"
    # ccc = '<pad> ' + ccc
    # ccc_output = '<start> pi:4a:7'
    output = ''
        
    output += (ccc_output)
    yield output
    occurrence = {}
    state = None
    for i in range(length): # only trained with ctx4096 (will be longer soon)
        if i == 0:
            out, state = model.forward(tokenizer.encode(ccc), state)
        else:
            out, state = model.forward([token], state)

        if MIDI_MODE: # seems only required for MIDI mode
            for n in occurrence:
                out[n] -= (0 + occurrence[n] * 0.5)
        
            out[0] += (i - 2000) / 500 # not too short, not too long
            out[127] -= 1 # avoid "t125"
            if piano_only:
                out[128:12416] -= 1e10
                out[13952:20096] -= 1e10
        token = pipeline.sample_logits(out, temperature=1.0, top_k=8, top_p=0.8)
        if token == EOS_ID: break
        
        if MIDI_MODE: # seems only required for MIDI mode
            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 += TOKEN_SEP + tokenizer.decode([token])
        yield output

    output += (' <end>')
    yield output