inference4.py

# RSICD caption generation script
import numpy as np
import torch
import torchvision.transforms as T
from decord import VideoReader, cpu
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer
import json
import os

import os
import random
from datasets import load_dataset
from PIL import Image

# 加载 RSICD 数据集
dataset = load_dataset("arampacha/rsicd")

# 定义输出目录
output_dirs = {
    "train": "train",
    "test": "test",
    "valid": "valid"
}

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)

def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform

def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio

def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

# If you have an 80G A100 GPU, you can put the entire model on a single GPU.
# Otherwise, you need to load a model using multiple GPUs, please refer to the `Multiple GPUs` section.
path = '/data/yyf/model/InternVL3-8B'
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    use_flash_attn=True,
    trust_remote_code=True).eval().cuda()
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)

generation_config = dict(max_new_tokens=1024, do_sample=False)

q = """
Task Description:
You are given a remote sensing image along with five incomplete captions for the image. Your task is to generate a comprehensive and accurate caption that fully describes the key features of the image. Use the provided captions as a reference, but ensure your final caption is complete and concise, capturing all major elements of the image.
Instructions:
1. Review the five provided captions.
2. Analyze the content of the image.
3. Generate a single, comprehensive caption that describes the key elements in the image in a clear and concise manner.
4. The caption should cover all important aspects of the image without being too long.
5. Avoid introductory phrases like "This is an image of..." or "This is a satellite image showing...". Focus on directly describing the image.
6. Ensure the caption is detailed enough to convey the essential information.
Here are the remote sensing image and the five incomplete captions:\n
"""

def process_and_save(dataset_split):
    with open("/data/xcl/dataSet/RSICD_1/descriptions_test_full.txt", 'w', encoding='utf-8') as f_out:
        for idx, example in enumerate(dataset_split, start=1):
            filename = f"{idx}.png"
            filepath = os.path.join(r"/data/xcl/dataSet/RSICD_1/test_png", filename)
            pixel_values = load_image(filepath).to(torch.bfloat16).cuda()
            caption = str(example["captions"])
            question = q + '<image>\n' + f"Captions: {caption}"
            response = model.chat(tokenizer, pixel_values, question, generation_config)
            response_single_line = response.strip().replace('\n', ' ').replace('\r', ' ')
            print(f'User: {question}\nAssistant: {response_single_line}')
            f_out.write(f"{idx} {response_single_line}\n")

# 执行处理
for split, out_dir in output_dirs.items():
    if split == "test":
        print(f"Processing {split} set ...")
        process_and_save(dataset[split])
print("All done!")