cmi / MTG /sft.py
nicolaus625's picture
Add files using upload-large-folder tool
ec23e22 verified
raw
history blame
2.48 kB
import json
from emotion import get_emotion
from genre import get_genre
from instrument import get_instrument
import os
from pydub import AudioSegment
import tqdm
EXCERPT_LENGTH = 30 * 1000 # 30 seconds in milliseconds
MIN_LENGTH = 5 * 1000 # 5 seconds in milliseconds
# Save to JSONL format
for split in ["test", "train", "valid"]:
output_file_path = f'MTG-Jamendo_{split}.jsonl' # Replace with the desired output path
emotion = get_emotion(_=split)
genre = get_genre(_=split)
instrument = get_instrument(_=split)
data_samples = emotion + genre + instrument
with open(output_file_path, 'w') as outfile:
for sample in tqdm.tqdm(data_samples):
# json.dump(sample, outfile)
audio_path = f"/home/intern-2024-02/dataset/MTG-Jamendo/MTG/audio-low/{sample['audioid'][:-4]}.low.mp3"
# Check if the audio file exists
if not os.path.exists(audio_path):
print(f"File not found: {audio_path}")
continue
# Load the audio file
audio = AudioSegment.from_file(audio_path)
# Process and split the audio into 30-second excerpts
for i in range(0, len(audio), EXCERPT_LENGTH):
end = i + EXCERPT_LENGTH
if end < len(audio):
excerpt = audio[i:end]
else:
excerpt = audio[i:]
# Discard short audio clips
if len(excerpt) < MIN_LENGTH:
break
end = len(audio)
# # Save the excerpt to the same directory with a new name
excerpt_path = f"/work/fast_data_yinghao/MTG/audio-low/{sample['audioid'][:-4]}.low_{i//EXCERPT_LENGTH}.wav"
# excerpt.export(excerpt_path, format="wav")
# Update the sample dictionary with the new audio path
new_sample = sample.copy()
new_sample["audioid"] = excerpt_path
# Dump the new sample to the JSONL file
json.dump(new_sample, outfile)
outfile.write('\n')
# print(excerpt_path)
# break
# Remove the original audio file
# os.remove(audio_path)
# outfile.write('\n')
outfile.close()
# print(f"Data successfully transformed and saved to {output_file_path}")