feature_combiner.py

import json
import random

class FeatureCombiner():
    def __init__(self, json_file, dataset=None) -> None:
        with open(json_file, "r") as f:
            data = json.load(f)
        
        self.data = data
        self.seg_data = [entry for entry in self.data if "annotations" in entry["image"]]
        self.img_data = [entry for entry in self.data if "images" in entry["image"]]
        
        self.img_data = [entry for entry in self.img_data if dataset in entry["dataset"]]
        self.seg_data = [entry for entry in self.seg_data if dataset in entry["dataset"]]

        self.dataset_field = dataset
        if self.dataset_field is None:
            self.feature_name = ["dataset", "location", "size", "trend_shape"]
        else:
            self.feature_name = ["location", "size", "trend_shape"]

    def combine(self, is_seg_only=True, debug=False):
        features_seg = []
        features_img = []
        # 选择seg_data的overview，这里假设每个entry中都有"overview"键
        seg_overview_entry = random.choice(self.seg_data)
        overview_seg = seg_overview_entry["overview"]
        overview_img = self._find_img_entry_by_seg_entry(seg_overview_entry)["overview"]

        features_seg.append(overview_seg)
        features_img.append(overview_img)
        if self.dataset_field != None:
            features_seg.append(self.dataset_field)
            features_img.append(self.dataset_field)
        else:
            raise "Not implement"
        
        if not debug:
            tmp_f_seg = []
            tmp_f_img = []
            # 根据is_seg_only选择数据集
            data = self.seg_data if is_seg_only else self.data
            for feature in self.feature_name:
                # 随机选择特征，这里假设每个entry中都有这些feature_name中定义的键
                entry_seg = random.choice(data)
                entry_img = self._find_img_entry_by_seg_entry(entry_seg)
                tmp_f_seg.append(entry_seg[feature])    
                tmp_f_img.append(entry_img[feature])    
            features_seg = features_seg + tmp_f_seg
            features_img = features_img + tmp_f_img
        
        caption_seg = ""
        for sentence in features_seg:
            # 整理sentence: 删除末尾的标点符号，去掉结尾多余的空格，然后在结尾添加分号，最后拼接到caption中
            sentence = sentence.rstrip(';.!?')  # 删除末尾的标点符号
            sentence = sentence.strip()  # 去掉结尾多余的空格
            sentence += ";"  # 在结尾添加分号
            caption_seg += sentence + " "  # 拼接到caption中，并添加一个空格以分隔句子
        
        # 删除caption最后的分号和空格
        caption_seg = caption_seg.rstrip('; ')

        caption_img = ""
        for sentence in features_img:
            # 整理sentence: 删除末尾的标点符号，去掉结尾多余的空格，然后在结尾添加分号，最后拼接到caption中
            sentence = sentence.rstrip(';.!?')  # 删除末尾的标点符号
            sentence = sentence.strip()  # 去掉结尾多余的空格
            sentence += ";"  # 在结尾添加分号
            caption_img += sentence + " "  # 拼接到caption中，并添加一个空格以分隔句子
        
        # 删除caption最后的分号和空格
        caption_img = caption_img.rstrip('; ')
        
        return caption_seg, caption_img
    
    def _find_img_entry_by_seg_entry(self, seg_entry):
        seg_path = seg_entry["image"]
        file_name = seg_path.split('/')[-1].split(".")[0]
        if "gt" in file_name:
            file_name = file_name.replace("gt", "img")
        
        for entry in self.img_data:
            if file_name+"." in entry["image"]:
                return entry
        
        raise "No entry find"
        
            
if __name__  == "__main__":
    json_file = r"/data/leiqin/dataset_curvilinear/dataset_curvilinear/crack/crack.json"
    fc = FeatureCombiner(json_file,"Crack500 dataset")
    for _ in range(9999):
        caption1, caption2 = fc.combine()
        print(caption1, caption2)