超越图像偏好——文本到图像生成中的丰富人工反馈

首先安装 rapidata 软件包，然后创建 RapidataClient()，这将用于创建并启动标注设置。

  pip install rapidata

from rapidata import RapidataClient, LabelingSelection, ValidationSelection

client = RapidataClient()

作为示例数据，我们将只使用数据集中的图像。请务必设置 streaming=True，因为下载整个数据集可能需要大量时间。

from datasets import load_dataset

ds = load_dataset("Rapidata/text-2-image-Rich-Human-Feedback", split="train", streaming=True)
ds = ds.select_columns(["image","prompt"])

由于我们使用流式传输，我们可以像这样提取提示并下载所需的图像

import os
tmp_folder = "demo_images"


# make folder if it doesn't exist
if not os.path.exists(tmp_folder):
   os.makedirs(tmp_folder)


prompts = []
image_paths = []
for i, row in enumerate(ds.take(10)):
   prompts.append(row["prompt"])
   # save image to disk
   save_path = os.path.join(tmp_folder, f"{i}.jpg")
   row["image"].save(save_path)
   image_paths.append(save_path)

# Alignment Example 
instruction = "How well does the image match the description?"
answer_options = [
        "1: Not at all",
        "2: A little",
        "3: Moderately",
        "4: Very well",
        "5: Perfectly"
    ]

order = client.order.create_classification_order(
    name="Alignment Example",
    instruction=instruction,
    answer_options=answer_options,
    datapoints=image_paths,
    contexts=prompts, # for alignment, prompts are required as context for the annotators.
    responses_per_datapoint=10,
    selections=[ValidationSelection("676199a5ef7af86285630ea6"), LabelingSelection(1)] # here we use a pre-defined validation set. See https://docs.rapidata.ai/improve_order_quality/ for details
)

order.run() # This starts the order. Follow the printed link to see progress.

# alignment heatmap
# Note that the selected images may not actually have severely misaligned elements, but this is just for demonstration purposes.

order = client.order.create_locate_order(
    name="Alignment Heatmap Example",
    instruction="What part of the image does not match with the description? Tap to select.",
    datapoints=image_paths,
    contexts=prompts, # for alignment, prompts are required as context for the annotators.
    responses_per_datapoint=10,
    selections=[ValidationSelection("67689e58026456ec851f51f8"), LabelingSelection(1)] # here we use a pre-defined validation set for alignment. See https://docs.rapidata.ai/improve_order_quality/ for details
)

order.run() # This starts the order. Follow the printed link to see progress.

# Select words example

from rapidata import LanguageFilter

select_words_prompts = [p + " [No_Mistake]" for p in prompts]
order = client.order.create_select_words_order(
    name="Select Words Example",
    instruction = "The image is based on the text below. Select mistakes, i.e., words that are not aligned with the image.",
    datapoints=image_paths,
    sentences=select_words_prompts, 
    responses_per_datapoint=10,
    filters=[LanguageFilter(["en"])], # here we add a filter to ensure only english speaking annotators are selected
    selections=[ValidationSelection("6761a86eef7af86285630ea8"), LabelingSelection(1)] # here we use a pre-defined validation set. See https://docs.rapidata.ai/improve_order_quality/ for details
)

order.run()