VocalSet/sft.py

import os
import random
import json
import pandas as pd
import hashlib
import glob


number_to_letter = {
    0: "A",
    1: "B",
    2: "C",
    3: "D"
}

class2id = {'belt': 0, 'breathy': 1, 'inhaled': 2, 'lip_trill': 3, 'spoken': 4, 'straight': 5, 'trill': 6, 'trillo': 7, 'vibrato': 8, 'vocal_fry': 9}
id2class = {v: k for k, v in class2id.items()}

PATH = "/work/fast_data_yinghao/VocalSet"
data_samples = []
for split in ["train", "valid", 
              "test"]:
    metadata = pd.read_csv(filepath_or_buffer=os.path.join(PATH, f'{split}_t.txt'), 
                                        names = ['audio_path'])
    for index in range(metadata.shape[0]):
        audio_path = metadata.iloc[index][0]
        label = audio_path.split('/')[0]
        audioid = f"{audio_path.split('/')[1]}"
        data_sample = {
            "instruction": "Please recognise the vocal technique in the given audio.",
            "input": f"<|SOA|>f'{audioid}'<|EOA|>",
            "output": label,
            "uuid": audio_path,
            "audioid":audio_path,
            "split": [split if split != "valid" else "dev"],
            "task_type": {"major": ["global_MIR"], "minor": ["vocal_technique_classification"]},
            "domain": "music",
            "source": "internet",
            "other": {"tag":"null"}
        }
        # print(instrument)
        data_samples.append(data_sample)
        # if index > 2:
        #     break
    
existed_uuid_list = set()
all_instruments = set(k for k,v in class2id.items())
for data_sample in data_samples:
    # change testset instruction to the choice format
    # if data_sample["split"][0] != "test":
    data_sample["instruction"] = data_sample["instruction"] + " Output from the following options: "
    for k,v in class2id.items():
        data_sample["instruction"] += f"{k}, "
    data_sample["instruction"] =  data_sample["instruction"][:-2] + ". "
    # else:
    #     correct_instrument = data_sample["output"]
    #     incorrect_instruments = list(all_instruments - set(correct_instrument))
    #     if len(incorrect_instruments) >= 3:
    #         choices = random.sample(incorrect_instruments, 3) + ["CORRECT:" + correct_instrument]
    #         random.shuffle(choices)
    #         for idx, choice in enumerate(choices):
    #             if choice.startswith("CORRECT:"):
    #                 choices[idx] = choice[8:]
    #                 data_sample["output"] = number_to_letter[idx]
    #         data_sample["input"] = f"{data_sample['input']}. Choose from: A.{choices[0]} B.{choices[1]} C.{choices[2]} D.{choices[3]} "
    
    # 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}"
    
# Save to JSONL format
for split in ["train", "dev", "test"]:
    if split == "dev":
        name = "valid"
    else:
        name = split
    with open(f"VocalSet_{name}.jsonl", 'w') as outfile:
        for sample in data_samples:
            if sample["split"][0] == split:
                json.dump(sample, outfile)
                outfile.write('\n')
    outfile.close()
# print(f"Data successfully transformed and saved to {output_file_path}")