scripts/make_ocr.py

import os
import os.path as osp
import json
import random
import transformers
from tqdm import tqdm
from multiprocessing import Pool
from functools import partial
from copy import deepcopy

from data_class import domNode, nodeBbox


MAX_TOTAL_LEN = 800
MAX_TOKEN_LEN = 50
MIN_TOKEN_LEN = 0
MAX_IoU = 0

NODE_SEPARATOR_TOKEN = '<node_separator>'

TEXT2BBOX_TEMPLATE = [
    "Where is the given content?",
    # Other template questions here...
]

BBOX2TEXT_TEMPLATE = [
    "Can you give me a description of the region in image?",
    # Other template questions here...
]


class NodewithOCR(domNode):
    def __init__(self, id, info, children, father=None, tokenizer=None, img_w=None, img_h=None, task=""):
        super().__init__(id, info, children, father, tokenizer, img_w=img_w, img_h=img_h, task="")
        self.struct_text: str = self.get_struct_text()

    def get_struct_text(self) -> str:
        if self.is_leaf():
            struct_text = self.info.text.strip()
            if self.is_img():
                if not struct_text:
                    return ""
                if self.info.type == 'img[svg]':
                    struct_text = '[svg]' + struct_text
                else:
                    struct_text = '[image]' + struct_text
            return struct_text
        
        children_text = []
        for child in self.children:
            child_text = child.struct_text
            if child_text:
                children_text.append(child.struct_text)
        return NODE_SEPARATOR_TOKEN.join(children_text)
    
    def get_context_text(self, root_text: str) -> str:
        self.context_text: str = ""

        cur_node = self
        for _ in range(3):
            if root_text.count(cur_node.struct_text) == 1:
                func = ''
                if self.info.func == 'type':
                    func += '(input box)'
                elif self.info.func == 'click':
                    func += '(clickable)'
                elif self.info.func == 'text':
                    func += '(pure text)'
                if cur_node == self:
                    context = ""
                else:
                    context = f'(context: "{cur_node.struct_text}")'
                self.context_text = self.struct_text + func + context
                return

            if not cur_node.father:
                return

            cur_node = cur_node.father

    # x1, y1, x2, y2
    def bbox2str(self) -> str:
        x1_rel = round(max(0, (self.info.bbox.x1 / IMG_W)) * 1000)
        y1_rel = round(max(0, (self.info.bbox.y1 / IMG_H)) * 1000)
        x2_rel = round(min(0.999, (self.info.bbox.x2 / IMG_W)) * 1000)
        y2_rel = round(min(0.999, (self.info.bbox.y2 / IMG_H)) * 1000)

        coords = [x1_rel, y1_rel, x2_rel, y2_rel]
        bbox_str = ["{:03}".format(coord) for coord in coords]
        bbox_str = "[" + ", ".join(bbox_str) + "]"
        
        return bbox_str
    
    # x1, y1, x2, y2  seeclick_format
    def bbox2str_2(self) -> str:
        x1_rel = round(max(0, (self.info.bbox.x1 / IMG_W)), 2)
        y1_rel = round(max(0, (self.info.bbox.y1 / IMG_H)), 2)
        x2_rel = round(min(0.999, (self.info.bbox.x2 / IMG_W)), 2)
        y2_rel = round(min(0.999, (self.info.bbox.y2 / IMG_H)), 2)

        coords = [x1_rel, y1_rel, x2_rel, y2_rel]
        bbox_str = ["{:.2f}".format(coord) for coord in coords]
        bbox_str = "(" + ",".join(bbox_str) + ")"
        
        return bbox_str
    
    # bbox center_x, center_y, w, h
    def bbox2str_ct(self) -> str:
        cx_rel = round(min(0.999, max(0, (self.info.bbox.cx / IMG_W))) * 1000)
        cy_rel = round(min(0.999, max(0, (self.info.bbox.cy / IMG_H))) * 1000)
        w_rel = round(min(0.999, max(0, (self.info.bbox.width / IMG_W))) * 1000)
        h_rel = round(min(0.999, max(0, (self.info.bbox.height / IMG_H))) * 1000)

        coords = [cx_rel, cy_rel, w_rel, h_rel]
        bbox_str = ["{:03}".format(coord) for coord in coords]
        bbox_str = "[" + ", ".join(bbox_str) + "]"
        
        return bbox_str
    
    # block idx, bbox center_x, center_y, w, h
    def bbox2str_ct_block(self, block_num_w=2, block_num_h=2) -> str:
        block_w = IMG_W // block_num_w
        block_h = IMG_H // block_num_h

        block_xi = max(0, min(IMG_W - 1, self.info.bbox.cx)) // block_w
        block_yi = max(0, min(IMG_H - 1, self.info.bbox.cy)) // block_h
        block_idx = block_yi * block_num_w + block_xi

        cx_block = self.info.bbox.cx % block_w
        cy_block = self.info.bbox.cy % block_h

        cx_rel = round(min(0.999, max(0, (cx_block / block_w))) * 1000)
        cy_rel = round(min(0.999, max(0, (cy_block / block_h))) * 1000)
        w_rel = round(min(0.999, max(0, (self.info.bbox.width / block_w))) * 1000)
        h_rel = round(min(0.999, max(0, (self.info.bbox.height / block_h))) * 1000)

        coords = [cx_rel, cy_rel, w_rel, h_rel]
        bbox_str = [str(block_idx)] + ["{:03}".format(coord) for coord in coords]
        bbox_str = "[" + ", ".join(bbox_str) + "]"
        
        return bbox_str
    
    # center point  seeclick_format
    def bbox2str_ct_2(self) -> str:
        cx_rel = round(min(0.999, max(0, (self.info.bbox.cx / IMG_W))), 2)
        cy_rel = round(min(0.999, max(0, (self.info.bbox.cy / IMG_H))), 2)

        coords = [cx_rel, cy_rel]
        bbox_str = ["{:.2f}".format(coord) for coord in coords]
        bbox_str = "(" + ",".join(bbox_str) + ")"
        
        return bbox_str


def init_context_text(node: NodewithOCR, root_text: str):
    node.get_context_text(root_text)
    for child in node.children:
        init_context_text(child, root_text)


tokenizer = transformers.AutoTokenizer.from_pretrained(
    "path/to/your/tokenizer",
    trust_remote_code=True,
)
tokenizer.add_tokens([NODE_SEPARATOR_TOKEN], special_tokens=True)


def get_str_token_len(str_: str) -> int:
    return len(tokenizer.encode(str_))


def collect_nodes(dom: NodewithOCR, node_list: list[NodewithOCR]) -> list:
    if dom.is_valid:
        node_list.append(dom)

    for child in dom.children:
        collect_nodes(child, node_list)

    return node_list


def select_node(node, node_text, node_bbox: nodeBbox, selected_node_list: list[domNode], total_len: int) -> tuple[bool, int]:
    if node.info.func == 'type':
        if not node.context_text:
            return False, total_len
    if not node_text:
        return False, total_len

    try:
        token_len = get_str_token_len(node_text)
    except:
        return False, total_len
    if token_len > MAX_TOKEN_LEN or token_len < MIN_TOKEN_LEN:
        return False, total_len
    
    total_len += token_len + 20

    if selected_node_list:
        selected_bbox_list = [_node.info.bbox for _node in selected_node_list]
        max_IoUs = node_bbox.get_max_IoU(selected_bbox_list)
        if max_IoUs > MAX_IoU:
            return False, total_len

    return True, total_len


def get_bbox_text(node_list: list[NodewithOCR], context=False, is_point=True, is_seeclick=False):  
    input_node_list = [_node for _node in node_list if _node.info.func == 'type']
    pointer_node_list = [_node for _node in node_list if _node.info.pointer]
    click_node_list = [_node for _node in node_list if not _node.info.pointer and _node.info.func == 'click']
    text_node_list = [_node for _node in node_list if _node.info.func == 'text']

    input_node_list_copy = deepcopy(input_node_list)
    pointer_node_list_copy = deepcopy(pointer_node_list)
    click_node_list_copy = deepcopy(click_node_list)
    text_node_list_copy = deepcopy(text_node_list)

    random.shuffle(input_node_list_copy)
    random.shuffle(pointer_node_list_copy)
    random.shuffle(click_node_list_copy)
    random.shuffle(text_node_list_copy)

    node_list_copy = input_node_list_copy + pointer_node_list_copy + click_node_list_copy + text_node_list_copy

    total_len = 0
    selected_node_list = []
    for node in node_list_copy:
        node_text = node.context_text if context else node.struct_text
        keep, total_len = select_node(node, node_text, node.info.bbox, selected_node_list, total_len)
        if total_len > MAX_TOTAL_LEN:
            break
        if keep:
            selected_node_list.append(node)
    random.shuffle(selected_node_list)

    bbox_str_list = []
    text_str_list = []
    for i, node in enumerate(selected_node_list):
        node_text = node.context_text if context else node.struct_text

        if not is_seeclick:
            bbox_str_list.append(f"{i+1}. {node.bbox2str_ct_block(block_num_w=BLOCK_NUM_W, block_num_h=BLOCK_NUM_H)}")
        else:
            if is_point:
                bbox_str_list.append(f"{i+1}. {node.bbox2str_ct_2()}")
            else:
                bbox_str_list.append(f"{i+1}. {node.bbox2str_2()}")

        text_str_list.append(f"{i+1}. {node_text}")
        
    bbox_str = '\n'.join(bbox_str_list)
    text_str = '\n'.join(text_str_list)

    return bbox_str, text_str


def make_bbox2text(node_list: list[NodewithOCR], img_path: str, is_point=True, is_seeclick=False):
    bbox_str, text_str = get_bbox_text(node_list, context=False, is_point=is_point, is_seeclick=is_seeclick)
    if not bbox_str or not text_str:
        return
    
    prompt = bbox_str + '\n' + random.choice(BBOX2TEXT_TEMPLATE)
    return {
        "image": img_path,
        "conversations": [
            {
                "from": "human",
                "value": f"<image>\n{prompt}"
            },
            {
                "from": "gpt",
                "value": text_str
            }
        ]
    }


def make_text2bbox(node_list: list[NodewithOCR], img_path: str, is_point=True, is_seeclick=False):
    bbox_str, text_str = get_bbox_text(node_list, context=True, is_point=is_point, is_seeclick=is_seeclick)
    if not bbox_str or not text_str:
        return
    
    prompt = text_str + '\n' + random.choice(TEXT2BBOX_TEMPLATE) + '(bbox: [block_index, cx, cy, w, h])'

    return {
        "image": img_path,
        "conversations": [
            {
                "from": "human",
                "value": f"<image>\n{prompt}"
            },
            {
                "from": "gpt",
                "value": bbox_str
            }
        ]
    }


def write_ann(ann_list, ann_path, mode='w', save_ratio=1.0):
    ann_dir = osp.dirname(ann_path)
    if not osp.exists(ann_dir):
        os.mkdir(ann_dir)

    if save_ratio == 1.0 or random.random() < save_ratio:
        with open(ann_path, mode, encoding='utf-8') as f:
            for i in range(len(ann_list)):
                ann_list[i]['id'] = i

                ann_str = json.dumps(ann_list[i], ensure_ascii=False)
                f.write(ann_str + '\n')


def single_proc(mode, data_dir, language, sub_dir):
    sub_path = osp.join(data_dir, sub_dir)
    screen_shot_dir = osp.join(sub_path, 'screenshot')
    dom_sub_dir = 'dom_svg' if language == 'zh' else 'dom_svg_en'
    dom_dir = osp.join(sub_path, dom_sub_dir)
    html_dir = osp.join(sub_path, 'html')

    html_path = osp.join(html_dir, 'html_0.html')
    if not osp.exists(html_path):
        return []
    html_content = open(html_path).readline().strip()
    if html_content in ['https://www.qq.com/', 'https://music.163.com/']:
        return []

    all_annotations = []
    for img_file in os.listdir(screen_shot_dir):
        if 'tmp' in img_file:
            continue
        file_ids = img_file.split('.')[0].split('_')[1]
        json_path = osp.join(dom_dir, f'dom_{file_ids}.json')
        img_path = osp.join(screen_shot_dir, img_file)
        if not osp.exists(json_path):
            continue

        dom_data = json.load(open(json_path))

        try:
            dom_node = NodewithOCR(tokenizer=tokenizer, img_w=IMG_W, img_h=IMG_H, task=mode, **dom_data)
        except ValueError as e:
            print(f"Json data error: {e}\n{json_path}")
            continue

        init_context_text(dom_node, dom_node.struct_text)

        node_list = collect_nodes(dom_node, [])
        if mode == 'bbox2text':
            ann = make_bbox2text(node_list, img_path, is_seeclick=False, is_point=False)
        elif mode == 'text2bbox':
            ann = make_text2bbox(node_list, img_path, is_seeclick=False, is_point=False)
        elif mode == 'text2bbox_seeclick':
            ann = make_text2bbox(node_list, img_path, is_seeclick=True, is_point=False)
        elif mode == 'bbox2text_seeclick':
            ann = make_bbox2text(node_list, img_path, is_seeclick=True, is_point=False)
        elif mode == 'text2point_seeclick':
            ann = make_text2bbox(node_list, img_path, is_seeclick=True, is_point=True)
        elif mode == 'point2text_seeclick':
            ann = make_bbox2text(node_list, img_path, is_seeclick=True, is_point=True)

        else:
            assert 0
        if ann is not None:
            all_annotations.append(ann)
    
    return all_annotations


def main_multi(data_dir, dst_path, mode, language='zh', save_ratio=1.0):
    pool = Pool(processes=16)

    sub_dir_list = os.listdir(data_dir)
    single_proc_partial = partial(single_proc, mode, data_dir, language)

    for res in tqdm(pool.imap_unordered(single_proc_partial, sub_dir_list), total=len(sub_dir_list)):
        write_ann(res, dst_path, mode='a', save_ratio=save_ratio)


if __name__ == '__main__':
    mode = 'text2bbox'
    # mode = 'bbox2text'
    
    # ZH
    BLOCK_NUM_W = 2
    BLOCK_NUM_H = 2
    IMG_W = 1120
    IMG_H = 1120

    data_dir = 'data_20240617'
    dst_path = f'xxx.jsonl'

    # EN
    BLOCK_NUM_W = 3
    BLOCK_NUM_H = 2
    IMG_W = 1920
    IMG_H = 1080

    data_dir = 'data_20240624'
    dst_path = f'xxx.jsonl'

    main_multi(data_dir, dst_path, mode, language='en')