将文本转换为音乐的 AI 实现贝特莱斯和奎恩之间的协作
我尝试了MubertAI将文本转换为音乐。
完成
请看一看。
披头士乐队和女王
贾斯汀·比伯和麦当娜
约翰·列侬和鲍勃·迪伦
肖邦和贝多芬
BTS 和黑色粉红色
猫王和埃尔顿·约翰
埃米内姆和泰勒·斯威夫特
涅磐和绿洲
流浪儿童和碧昂斯
埃德·希兰和查理·普斯
尝试
环境
- Python
- Google Colab
导入和安装
import subprocess, time
print("Setting up environment... ")
start_time = time.time()
all_process = [
['pip', 'install', 'torch==1.12.1+cu113', 'torchvision==0.13.1+cu113', '--extra-index-url', 'https:// download.pytorch.org/whl/cu113'],
['pip', 'install', '-U', 'sentence-transformers'],
['pip', 'install', 'httpx'],
]
for process in all_process:
running = subprocess.run(process,stdout=subprocess. PIPE).stdout.decode('utf-8')
end_time = time.time()
print(f"Environment set up in {end_time-start_time:.0f} seconds")API 设置
import numpy as np
from sentence_transformers import SentenceTransformer
minilm = SentenceTransformer('all-MiniLM-L6-v2')
mubert_tags_string = 'tribal,action,kids,neo-classic,run 130,pumped,jazz / funk,ethnic,dubtechno,reggae,acid jazz,liquidfunk,funk,witch house,tech house,underground, artists,mystical,disco,sensorium,r&b,agender,psychedelic trance / psytrance,peaceful,run 140,piano,run 160,setting,meditation,christmas,ambient,horror,cinematic, electro house,idm,bass,minimal,underscore,drums,glitchy,beautiful,technology,tribal house,country pop,jazz & funk,documentary,space,classical,valentines,chillstep, experimental,trap,new jack swing,drama,post-rock,tense,corporate,neutral,happy,analog,funky,spiritual,sberzvuk special,chill hop,dramatic,catchy,holidays,fitness 90, optimistic,orchestra,acid techno,energizing,romantic,minimal house,breaks,hyper pop,warm up,dreamy,dark,urban,microfunk,dub,nu disco,vogue,keys,hardcore,aggressive,indie, electro funk,beauty,relaxing,trance,pop,hiphop,soft,acoustic,chillrave / ethno-house,deep techno,angry,dance,fun,dubstep,tropical,latin pop,heroic,world music,inspirational ,uplifting,atmosphere,art,epic,advertising,chillout,scary,spooky,slow ballad,saxophone,summer,erotic,jazzy,energy 100,kara mar,xmas,atmospheric,indie pop,hip-hop,yoga, reggaeton,lounge,travel,running,folk,chillrave & ethno-house,detective,darkambient,chill,fantasy,minimal techno,special,night,tropical house,downtempo,lullaby, meditative,upbeat,glitch hop,fitness,neurofunk,sexual,indie rock,future pop,jazz,cyberpunk,melancholic,happy hardcore,family / kids,synths,electric guitar,comedy,psychedelic trance & psytrance,edm,psychedelic rock,calm,zen,bells,podcast,melodic house,ethnic percussion,nature,heavy,bassline,indie dance,techno,drumnbass,synth pop,vaporwave,sad ,8-bit,chillgressive,deep,orchestral,futuristic,hardtechno,nostalgic,big room,sci-fi,tutorial,joyful,pads,minimal 170,drill,ethnic 108,amusing,sleepy ambient,psychill, italo disco,lofi,house,acoustic guitar,bassline house,rock,k-pop,synthwave,deep house,electronica,gabber,nightlife,sport & fitness,road trip,celebration,electro,disco house,electronic'
mubert_tags = np.array(mubert_tags_string.split(','))
mubert_tags_embeddings = minilm.encode(mubert_tags)
from IPython.display import Audio, display
import httpx
import json
def get_track_by_tags(tags, pat, duration, maxit=20, autoplay=False, loop=False):
if loop:
mode = "loop"
else:
mode = "track"
r = httpx.post('https://api-b2b.mubert.com/v2/RecordTrackTTM',
json={
"method":"RecordTrackTTM",
"params": {
"pat": pat,
"duration": duration,
"tags": tags,
"mode": mode
}
})
rdata = json.loads(r.text)
assert rdata['status'] == 1, rdata['error']['text']
trackurl = rdata['data']['tasks'][0]['download_link']
print('Generating track ', end='')
for i in range(maxit):
r = httpx.get(trackurl)
if r.status_code == 200:
display(Audio(trackurl, autoplay=autoplay))
break
time.sleep(1)
print('.', end='')
def find_similar(em, embeddings, method='cosine'):
scores = []
for ref in embeddings:
if method == 'cosine':
scores.append(1 - np.dot(ref, em)/(np.linalg.norm(ref)*np.linalg.norm(em)))
if method == 'norm':
scores.append(np.linalg.norm(ref - em))
return np.array(scores), np.argsort(scores)
def get_tags_for_prompts(prompts, top_n=3, debug=False):
prompts_embeddings = minilm.encode(prompts)
ret = []
for i, pe in enumerate(prompts_embeddings):
scores, idxs = find_similar(pe, mubert_tags_embeddings)
top_tags = mubert_tags[idxs[:top_n]]
top_prob = 1 - scores[idxs[:top_n]]
if debug:
print(f"Prompt: {prompts[i]}nTags: {', '.join(top_tags)}nScores: {top_prob}nnn")
ret.append((prompts[i], list(top_tags)))
return ret点击 API（*请更改电子邮件）
email = "deeprecommend@gmail.com" #@param {type:"string"}
r = httpx.post('https://api-b2b.mubert.com/v2/GetServiceAccess',
json={
"method":"GetServiceAccess",
"params": {
"email": email,
"license":"ttmmubertlicense#f0acYBenRcfeFpNT4wpYGaTQIyDI4mJGv5MfIhBFz97NXDwDNFHmMRsBSzmGsJwbTpP1A6i07AXcIeAHo5",
"token":"4951f6428e83172a4f39de05d5b3ab10d58560b8",
"mode": "loop"
}
})
rdata = json.loads(r.text)
assert rdata['status'] == 1, "probably incorrect e-mail"
pat = rdata['data']['pat']
print(f'Got token: {pat}')生成
prompt = 'Beatles with Queen' #@param {type:"string"}
duration = 30 #@param {type:"number"}
loop = False #@param {type:"boolean"}
def generate_track_by_prompt(prompt, duration, loop=False):
_, tags = get_tags_for_prompts([prompt,])[0]
try:
get_track_by_tags(tags, pat, duration, autoplay=True, loop=loop)
except Exception as e:
print(str(e))
print('n')
generate_track_by_prompt(prompt, duration, loop)结论
我想与大数据贝叶斯量子神经网络算法合作。 — Jin Sugimoto (@Jin__Sugimoto) October 21, 2022
AI就像中国产品，
现在只生产劣质产品，但
也许它们更便宜、更快、
更漂亮。 — Jin Sugimoto (@Jin__Sugimoto) October 21, 2022
DeepRecommend
深度网络工作，旨在成为一个巨大的技术经济区