data/hyper/enmap-pixtral-12B.py

import os
import torch
import dataset4eo as eodata
import jsonlines
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

from mistral_inference.transformer import Transformer #type:ignore
from mistral_inference.generate import generate #type:ignore

from mistral_common.tokens.tokenizers.mistral import MistralTokenizer #type:ignore
from mistral_common.protocol.instruct.messages import UserMessage, TextChunk, ImageChunk #type:ignore
from mistral_common.protocol.instruct.request import ChatCompletionRequest #type:ignore


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

    current_classes, current_percents = get_significant_classes(label)
    unknownId = -1
    current_class_names = {i:class_names[ind] for i,ind in enumerate(current_classes) if ind!=unknownId}
    current_percents = {i:percent for i,percent in enumerate(current_percents.values())}

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

    prompt = generate_prompt_for_segmentation(current_color_names, current_class_names, current_percents)
    label_remap = {i:ind for i,ind in enumerate(current_classes)}
    image = generate_color_coded_segmentation_map(label, class_colors, label_remap=label_remap)

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

    completion_request = ChatCompletionRequest(messages=[UserMessage(content=[ImageChunk(image=image), TextChunk(text=prompt)])])

    encoded = tokenizer.encode_chat_completion(completion_request)

    images = encoded.images
    tokens = encoded.tokens

    out_tokens, _ = generate([tokens], model, images=[images], max_tokens=256, temperature=0.35, eos_id=tokenizer.instruct_tokenizer.tokenizer.eos_id)
    result = tokenizer.decode(out_tokens[0])
    return result



if __name__=="__main__":
    import tqdm #type: ignore
    import json
    import argparse

    parser = argparse.ArgumentParser(description="Process some integers.")
    
    # Add the --part argument
    parser.add_argument('--part', type=int, required=True, help="Specify the part number as an integer.")
    
    # Parse the command-line arguments
    args = parser.parse_args()
    
    # Access the --part argument
    part = args.part

    # Generate partitions
    n_total = 15000
    n_parts = 10
    step = n_total // n_parts

    partitions = {
        i: {"start": (i - 1) * step, "end": i * step} for i in range(1, n_parts + 1)
    }

    partitions[1]['start']=1254
    index_dict = partitions[part]
    # load model
    mistral_models_path = "Pixtral-12B"
    tokenizer = MistralTokenizer.from_file(f"{mistral_models_path}/tekken.json")
    model = Transformer.from_folder(mistral_models_path)

    # load data
    dataset = eodata.StreamingDataset(input_dir="optimized_enmap_nlcd_dataset", num_channels=202, channels_to_select=[0,1,2], shuffle=False, drop_last=True)
    class_names = {
        11: "Open Water",
        12: "Perennial Ice/Snow",
        21: "Developed, Open Space",
        22: "Developed, Low Intensity",
        23: "Developed, Medium Intensity",
        24: "Developed, High Intensity",
        31: "Barren Land (Rock/Sand/Clay)",
        41: "Deciduous Forest",
        42: "Evergreen Forest",
        43: "Mixed Forest",
        51: "Dwarf Scrub",
        52: "Shrub/Scrub",
        71: "Grassland/Herbaceous",
        72: "Sedge/Herbaceous",
        73: "Lichens",
        74: "Moss",
        81: "Pasture/Hay",
        82: "Cultivated Crops",
        90: "Woody Wetlands",
        95: "Emergent Herbaceous Wetlands"
    }

    class_names = {}

    filename = f"hyper_id_text_nlcd_part{part}.jsonl"
    for id in tqdm.tqdm(range(index_dict['start'], index_dict['end'])):
        caption = get_caption_from_id(model, dataset, class_names, id)
        record = {"id": id, "caption": caption}  # Create a structured record
        with jsonlines.open(filename, mode='a') as writer:
            writer.write(record)  # Write one record per line