utils/expand_dataset.py

import os
import shutil
import json
import random

# prompt in llava
prompt_for_image = [
    "Describe the image concisely.",
    "Provide a brief description of the given image.",
    "Offer a succinct explanation of the picture presented.",
    "Summarize the visual content of the image."
    "Give a short and clear explanation of the subsequent image.",
    "Share a concise interpretation of the image provided.",
    "Present a compact description of the photo's key features.",
    "Relay a brief, clear account of the picture shown.",
    "Render a clear and concise summary of the photo.",
    "Write a terse but informative summary of the picture.",
    "Create a compact narrative representing the image presented.",
]

# path to the generated images
source_folder = "/mnt/petrelfs/zhuchenglin/diffusion/images_large1"
# path to the llava training images, which has the initial 660 subfolders
target_folder = "/mnt/petrelfs/zhuchenglin/LLaVA/playground/data/LLaVA-Pretrain/images"
# path to the llava training annotations folder
target_anno_folder = "/mnt/petrelfs/zhuchenglin/LLaVA/playground/data/LLaVA-Pretrain"
# path to the COCO annotations file
annotations_coco_path = (
    "/mnt/petrelfs/zhuchenglin/diffusion/coco/annotations/captions_train2017.json"
)
with open(annotations_coco_path, "r") as f:
    annotations = json.load(f)

new_annotations = []
for index, annotation in enumerate(annotations["annotations"][:500000]):
    print(index)
    # 660 is the starting index
    folder_index = 900 + (index // 10000)
    target_subfolder = f"{folder_index:05d}"

    # format of the image name: 00000xxxx.jpg
    target_image_name = f"{folder_index:05d}{index % 10000:04d}.jpg"
    target_image_path = os.path.join(target_folder, target_subfolder, target_image_name)
    if not os.path.exists(os.path.join(target_folder, target_subfolder)):
        os.makedirs(os.path.join(target_folder, target_subfolder))

    # the default name of generated images is index.jpg
    source_image_path = os.path.join(source_folder, f"{index}.jpg")
    if os.path.exists(source_image_path):
        shutil.copy(source_image_path, target_image_path)

    random_prompt = random.choice(prompt_for_image)
    new_annotation = {
        "id": f"{target_subfolder}{index % 10000:04d}",
        "image": f"{target_subfolder}/{target_image_name}",
        "conversations": [
            {"from": "human", "value": f"{random_prompt}\n<image>"},
            {"from": "gpt", "value": annotation["caption"]},
        ],
    }
    new_annotations.append(new_annotation)

json_file_path = os.path.join(target_anno_folder, "coco_annotations_500k.json")
with open(json_file_path, "w") as json_file:
    json.dump(new_annotations, json_file, indent=4)