data/elevation/batch_inference.py

import torch
from PIL import Image
from transformers import AutoModelForCausalLM
import dataset4eo as eodata
import json
import numpy as np
import pdb
from prompt_utils import generate_prompt_for_segmentation,\
    generate_color_coded_segmentation_map, get_significant_classes,\
    get_tableau_colors, resize_and_encode_image

from openai import OpenAI


def get_caption_from_id(dataset, class_names, id):
    class_colors = get_tableau_colors()  # {'blue': (31, 119, 180), 'orange': (255, 127, 14)}
    sample = dataset[id]
    label = sample["label"]
    # save image for debuging
    #rgb = (sample["image"]*255).astype(np.uint8)
    #rgb = Image.fromarray(rgb)
    #rgb.save(f"img_{id}.png")

    current_classes = get_significant_classes(label)
    unknownId = 12
    current_class_names = {ind:class_names[str(ind)] for ind in current_classes if ind!=unknownId}

    color_names = list(class_colors.keys())
    current_color_names = {ind:color_names[ind] for ind in current_classes}

    prompt = generate_prompt_for_segmentation(current_color_names, current_class_names)
    image = generate_color_coded_segmentation_map(label, class_colors)

    #print(prompt)
    #image.save(f"label_color_{id}.png")

    base_img = resize_and_encode_image(image)

    messages = [
        {"role": "user", "content": [
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{base_img}"}},
            {"type": "text", "text": prompt}
        ]}
    ]
    chat_completion = openai.chat.completions.create(
        model="meta-llama/Llama-3.2-90B-Vision-Instruct",
        messages=messages,
    )
    return chat_completion.choices[0].message.content




if __name__=="__main__":
    import tqdm #type: ignore
    import json
    # load model
    openai = OpenAI(
        api_key="uTu2kzPb6L08aXsmwwRI462UExeUtTBZ",
        base_url="https://api.deepinfra.com/v1/openai",
    )

    # load data
    dataset = eodata.StreamingDataset(input_dir="optimized_flair2_test", num_channels=5, channels_to_select=[0,1,2], shuffle=True, drop_last=True)
    meta_data = json.load(open("optimized_flair2_test/metadata.json",'r'))
    class_names = meta_data["attributes"]["class"]

    data = {}

    for id in tqdm.tqdm(range(len(dataset))):
        caption = get_caption_from_id(dataset, class_names, id)
        data[id] = caption
        if id%10==0:
            with open(f"rgb_captions_{id}.json", "w") as json_file:
                json.dump(data, json_file)

    # Write to JSON file
    with open("rgb_captions.json", "w") as json_file:
        json.dump(data, json_file)