cmi / music4all /sft.py
nicolaus625's picture
Add files using upload-large-folder tool
bb01eca verified
raw
history blame
5.01 kB
import hashlib
import json
import tqdm
import os
import pandas as pd
PATH = "/work/fast_data_yinghao/music4all"
pd_list = pd.read_csv(f"{PATH}/id_genres.csv", sep="\t")
genre_dict = pd_list.set_index('id')['genres'].to_dict()
# genres = set([j for i in genre_dict.values() for j in i.split(",")])
pd_list = pd.read_csv(f"{PATH}/id_lang.csv", sep="\t")
lang_dict = pd_list.set_index('id')['lang'].to_dict()
# langs = set([j for i in lang_dict.values() for j in i.split(",")])
pd_list = pd.read_csv(f"{PATH}/id_tags.csv", sep="\t")
tag_dict = pd_list.set_index('id')['tags'].to_dict()
# tags = set([j for i in tag_dict.values() for j in i.split(",")])
test_jsons = json.load(open(f"{PATH}/SongInterpretation/dataset_test.json","r"))
train_jsons = json.load(open(f"{PATH}/SongInterpretation/dataset_not_negative_256_clean.json","r"))
train_list = [i["music4all_id"] for i in train_jsons]
# train_dict = {i["music4all_id"]:i["comment"] for i in train_jsons} 1 id may have multiple comments
test_list = [i["music4all_id"] for i in test_jsons]
# test_dict = {i["music4all_id"]:i["comment"] for i in test_jsons}
existed_uuid_list = set()
def get_sample(id, instruction, output, task, split="train"):
data_sample = {
"instruction": instruction,
"input": f"<|SOA|>{id}.wav<|EOA|>",
"output": output,
"uuid": "",
"audioid": f"{id}.wav",
"split": [split],
"task_type": {"major": ["global_MIR"], "minor": [task]},
"domain": "music",
"source": "Music4All",
"other": {}
}
# change uuid
uuid_string = f"{data_sample['instruction']}#{data_sample['input']}#{data_sample['output']}"
unique_id = hashlib.md5(uuid_string.encode()).hexdigest()[:16] #只取前16位
if unique_id in existed_uuid_list:
sha1_hash = hashlib.sha1(uuid_string.encode()).hexdigest()[:16] # 为了相加的时候位数对应上 # 将 MD5 和 SHA1 结果相加,并计算新的 MD5 作为最终的 UUID
unique_id = hashlib.md5((unique_id + sha1_hash).encode()).hexdigest()[:16]
existed_uuid_list.add(unique_id)
data_sample["uuid"] = f"{unique_id}"
return data_sample
genre_samples, lang_samples, tag_samples = [], [], []
comment_train_samples, comment_test_samples = [], []
count = 0
for id in tqdm.tqdm(genre_dict.keys()):
if id in test_list:
continue
audio_path = os.path.join(f"{PATH}", f"audios/{id}.wav")
data_sample = get_sample(id,
"What is the genre of this music?",
genre_dict[id],
"genre_classification")
genre_samples.append(data_sample)
if count < 10000:
data_sample = get_sample(id,
"Which language from the following list is this music? List: ['zh-cn', 'de', 'sw', 'el', 'en', 'cy', 'hu', 'ar', 'so', 'lt', 'ja', 'ru', 'es', 'fr', 'sk', 'bg', 'et', 'th', 'sq', 'INTRUMENTAL', 'lv', 'pa', 'cs', 'no', 'hi', 'ca', 'pt', 'ko', 'nl', 'fa', 'sv', 'tr', 'sl', 'bn', 'pl', 'uk', 'id', 'he', 'af', 'ro', 'hr', 'it', 'vi', 'fi', 'tl', 'da']",
lang_dict[id],
"language_detection")
lang_samples.append(data_sample)
if lang_dict[id] == "en":
count += 1
data_sample = get_sample(id,
"What are the tags of this music?",
tag_dict[id],
"music_tagging")
tag_samples.append(data_sample)
# break
comment_test_samples = [get_sample(data["music4all_id"],
"You are the user of Spotify, please give a commments on the interpretation of the lyrics.",
data["comment"],
"lyrics_interpretation",
split="test")
for data in test_jsons]
comment_train_samples = [get_sample(data["music4all_id"],
"You are the user of Spotify, please give a commments on the interpretation of the lyrics.",
data["comment"],
"lyrics_interpretation")
for data in train_jsons]
print("genre_samples:", len(genre_samples))
print("lang_samples:", len(lang_samples))
print("tag_samples:", len(tag_samples))
print("comment_train_samples:", len(comment_train_samples))
print("comment_test_samples:", len(comment_test_samples))
# Save to JSONL format
output_file_path = f'{PATH}/Music4all_train.jsonl' # Replace with the desired output path
with open(output_file_path, 'w') as outfile:
# for sample in data_samples:
json.dump( comment_train_samples, outfile)
# genre_samples + tag_samples + lang_samples +
# outfile.write('\n')
outfile.close()
output_file_path = f'{PATH}/Music4all_test.jsonl'
with open(output_file_path, 'w') as outfile:
json.dump(comment_test_samples, outfile)
outfile.close()