aitw/prepare_trajectory_grounding.py

import json
from copy import deepcopy
import os
from pathlib import Path
import re
from PIL import Image
from tqdm import tqdm
import multiprocessing

import sys
sys.path.append('./google-research')
from android_in_the_wild.action_type import ActionType
from android_in_the_wild.visualization_utils import is_tap_action
import numpy as np

def is_tap(step):
    return step["results/action_type"][0] == ActionType.DUAL_POINT and is_tap_action(np.array(step["results/yx_touch"]), np.array(step["results/yx_lift"]))

def encode_action(action_json):
    """
    Encode different types of actions into human-readable descriptions.
    
    Args:
        action_json (dict): A dictionary containing action details
    
    Returns:
        str: A human-readable description of the action
    """
    action_type = action_json["results/action_type"][0]
    
    if is_tap(action_json):
        return "tap on the screen"

    elif action_type == ActionType.DUAL_POINT:
        # Check scroll direction by comparing y-values
        start_y, end_y = action_json["results/yx_touch"][1], action_json["results/yx_lift"][1]
        start_x, end_x = action_json["results/yx_touch"][0], action_json["results/yx_lift"][0]
        # first decide scroll vertically or horizontally
        if abs(start_y - end_y) > abs(start_x - end_x):
            if start_y < end_y:
                return "scroll down"
            else:
                return "scroll up"
        else:
            if start_x < end_x:
                return "scroll right"
            else:
                return "scroll left"
    
    elif action_type == ActionType.TYPE:
        text_to_type = action_json["results/type_action"][0]
        return f'TYPE "{text_to_type}"'
    
    elif action_type == ActionType.PRESS_BACK:
        return "go to the previous screen"
    
    elif action_type == ActionType.PRESS_HOME:
        return "go to the home screen"
    
    elif action_type == ActionType.PRESS_ENTER:
        return "press the enter key"
    
    elif action_type == ActionType.STATUS_TASK_COMPLETE:
        return "task completed"

    elif action_type == ActionType.STATUS_TASK_IMPOSSIBLE:
        return "task impossible"

    else:
        raise ValueError(f"Unknown action type: {action_type}")

def resize_image(image, scale=0.75):
    """
    Resize image to have its shorter edge equal to 720 pixels while maintaining aspect ratio.
    
    Args:
    image: PIL Image object
    
    Returns:
    Resized PIL Image
    """
    # Get current dimensions
    width, height = image.size
    
    # Calculate new dimensions
    new_width = int(width * scale)
    new_height = int(height * scale)
    
    # Resize image
    resized_image = image.resize((new_width, new_height), Image.LANCZOS)
    return resized_image

def merge_convs(conversations):
    """
    Merge all successive 'human' conversations comprehensively.
    
    Args:
        conversations (list): List of conversation dictionaries
    
    Returns:
        list: Processed conversations with all successive human messages merged
    
    Raises:
        ValueError: If input is not a list or contains invalid conversation dictionaries
    """
    # Validate input
    if not isinstance(conversations, list):
        raise ValueError("Input must be a list of conversation dictionaries")
    
    # Validate each conversation dictionary structure
    for conv in conversations:
        if not isinstance(conv, dict):
            raise ValueError("Each conversation must be a dictionary")
        if 'from' not in conv or 'value' not in conv:
            raise ValueError("Each conversation must have 'from' and 'value' keys")
    
    processed_conversations = []
    i = 0
    while i < len(conversations):
        current_conv = conversations[i]
        
        # If current conversation is 'human', start merging
        if current_conv['from'] == 'human':
            # Collect all successive human conversations
            merged_value = current_conv['value']
            j = i + 1
            while j < len(conversations) and conversations[j]['from'] == 'human':
                merged_value += '\n\n' + conversations[j]['value']
                j += 1
            
            # Update current conversation with merged value
            current_conv['value'] = merged_value
            
            # Move index to last non-human conversation
            i = j
        else:
            # For non-human conversations, just add to processed list
            i += 1
        
        processed_conversations.append(current_conv)
    
    return processed_conversations


def parse_reasoning(input_string):
    input_string = input_string.strip()
    if not input_string.endswith("```"):
        input_string += "```"
    # Regex pattern to match texts between ```A```, ```B```, and ```C```
    pattern = r'```([ABC])\n(.*?)```'
    
    # Find all matches
    matches = re.findall(pattern, input_string, re.DOTALL)
    
    # Create a dictionary to store parsed texts
    parsed_texts = []
    
    # Populate the dictionary
    for _, text in matches:
        parsed_texts.append(text.strip())

    if len(parsed_texts) != 3:
        # print(input_string)
        return None, None, None

    caption, reasoning, instruction = parsed_texts
    
    return caption, instruction.replace("Task: ", ""), reasoning

grounding_step_prompt = "<|img|>Step {step_idx}. Given a GUI image, what are the relative (0-1000) pixel point coordinates for the element corresponding to the following instruction or description: {instruction}"
grounding_step_ans = "```\n{point_str}\n```"
act_step_prompt = "<|img|>Step {step_idx}. Instruction: {prev_instruction}"
act_step_ans = "The agent's action: {prev_action}"
user_start_prompt = "The agent is performing the ultimate task: {ultimate_task}."
user_history_instr_prompt = "History of the agent's steps:\n{history_list}."

resize_ratios_per_window_size = {
    1: 0.5,
    2: 0.5,
    3: 0.5,
}

def process_android_episodes(data, window_size=2):
    """
    Process Android episodes and extract steps with click or long_press actions.
    
    Args:
    data (list): List of episode dictionaries
    window_size (int, optional): Number of recent image-included conversations to include. 
                                 Defaults to 3 (current image + 2 previous image-included steps).
    
    Returns:
    dict: Dictionary with episode_id as key and list of filtered steps as value
    """
    instructions = []
    for episode in data:
        episode_id = episode["episode_id"]
        
        for i, step in enumerate(episode["steps"]):
            is_grounding = step["is_grounding"]

            if not is_grounding:
                continue

            if window_size > 0 and i == 0: # skip the first step if window_size > 0
                continue
            
            convs = [
                {
                    "from": "human",
                    "value": user_start_prompt.format(
                        ultimate_task=episode["goal_info"][0]
                    ),
                },
            ]

            cur_img_list = [Path(step["image_path"]).resolve()]

            if window_size > 0:
                window_steps = episode["steps"][i-window_size:i] if i >= window_size else episode["steps"][:i]

                if i > window_size: # has more history steps larger than window_size
                    convs.append(
                        {
                            "from": "human",
                            "value": user_history_instr_prompt.format(
                                history_list="\n".join(
                                    [
                                        f"\t{j+1}. " + prev_step["step_instruction"]
                                        for j, prev_step in enumerate(episode["steps"][:i-window_size])
                                    ]
                                )
                            ),
                        },
                    )

                convs.append(
                    {
                        "from": "human",
                        "value": "The recent steps with the GUI images are as follows:\n",
                    }
                )

                for j, win_step_i in enumerate(window_steps):
                    if win_step_i["is_grounding"]:
                        convs.append(
                            {
                                "from": "human",
                                "value": grounding_step_prompt.format(
                                    instruction=win_step_i["step_instruction"], step_idx=i+1-(len(window_steps)-j)
                                ),
                            }
                        )
                        convs.append(
                            {
                                "from": "gpt",
                                "value": grounding_step_ans.format(point_str=f"({win_step_i['coord_norm'][0]}, {win_step_i['coord_norm'][1]})"),
                            }
                        )
                    else:
                        convs.append(
                            {
                                "from": "human",
                                "value": act_step_prompt.format(
                                    prev_instruction=encode_action(win_step_i), step_idx=i+1-(len(window_steps)-j)
                                ),
                            }
                        )
                        convs.append(
                            {
                                "from": "human",
                                "value": act_step_ans.format(
                                    prev_action=encode_action(win_step_i)
                                ),
                            }
                        )
                win_img_list = [
                    Path(win_step["image_path"]).resolve() for win_step in window_steps
                ]

                if not all([img_path.exists() for img_path in img_list]):
                    print(f"Image not found for episode {episode_id}, step {i+1}. Skipping...")
                    continue
                
                has_img_broken = False
                for img_path in img_list:
                    try:
                        Image.open(str(img_path))
                    except Exception as e:
                        print(f"Error opening image {img_path}: {e}")
                        has_img_broken = True
                        break
                if has_img_broken:
                    print(f"Image broken for episode {episode_id}, step {i+1}. Skipping...")
                    continue
                
                resize_scale = resize_ratios_per_window_size[window_size]
                win_img_list_resized = []
                try:
                    for img_path in win_img_list:
                        new_save_name = img_path.stem + f"_{resize_scale}x" + img_path.suffix
                        new_save_dir = img_path.parent.parent / f"images_resized"
                        new_save_dir.mkdir(parents=True, exist_ok=True)
                        new_save_path = new_save_dir / new_save_name
                        if new_save_path.exists():
                            win_img_list_resized.append(new_save_path)
                            continue
                        win_img = Image.open(str(img_path))
                        win_img = resize_image(win_img, scale=resize_scale)
                        win_img.save(str(new_save_path))
                        win_img_list_resized.append(new_save_path)
                except Exception as e:
                    print(f"Error resizing image: {e}: {win_img_list}")
                    continue

            else:
                convs.append(
                    {
                        "from": "human",
                        "value": user_history_instr_prompt.format(
                            history_list="\n".join(
                                [
                                    f"\t{j+1}. " + prev_step["step_instruction"]
                                    for j, prev_step in enumerate(episode["steps"][:i-window_size])
                                ]
                            )
                        ),
                    },
                )

            if window_size > 0:
                img_list = win_img_list_resized + cur_img_list
            else:
                img_list = cur_img_list

            has_img_broken = False
            for img_path in img_list:
                try:
                    Image.open(str(img_path))
                except Exception as e:
                    print(f"Error opening image {img_path}: {e}")
                    has_img_broken = True
                    break
            if has_img_broken:
                print(f"Image broken for episode {episode_id}, step {i+1}. Skipping...")
                continue

            # Current step details
            convs.append(
                {
                    "from": "human",
                    "value": grounding_step_prompt.format(instruction=step["step_instruction"], step_idx=i+1),
                }
            )
            convs.append(
                {
                    "from": "gpt",
                    "value": grounding_step_ans.format(point_str=f"({step['coord_norm'][0]}, {step['coord_norm'][1]})"),
                }
            )

            convs = merge_convs(convs)

            instructions.append(
                {
                    "image": [str(img_path) for img_path in img_list],
                    "conversations": convs,
                }
            )
    
    return instructions

# Example usage
if __name__ == "__main__":
    dataset_directories = {
        'general': './general_episodes_with_grounding_reasoning',
        'google_apps': './google_apps_episodes_with_grounding_reasoning',
        'install': './install_episodes_with_grounding_reasoning_valid',
        'web_shopping': './web_shopping_episodes_with_grounding_reasoning',
    }

    episode_data_list = []
    for dataset_name, directory_path in dataset_directories.items():
        episode_data_list.extend(list(Path(directory_path).glob("*.json")))
    episode_data_list = [json.load(open(str(file_path), "r", encoding="utf-8")) for file_path in episode_data_list]

    episode_data_list_new = []
    for episode_data in tqdm(episode_data_list, desc="Parsing fields..."):
        for step in episode_data:
            step["image_path"] = step["image_path"][0] if isinstance(step["image_path"], list) else step["image_path"]
            if "grounding_reasoning" not in step or not step["grounding_reasoning"]:
                step["step_instruction"] = encode_action(step)
                step["is_grounding"] = False
                continue

            caption, instruction, reasoning = parse_reasoning(step["grounding_reasoning"])
            step["step_instruction"] = instruction if instruction else encode_action(step)
            step["caption"] = caption
            step["reasoning"] = reasoning
            step["is_grounding"] = not(not instruction)
            step["coord_norm"] = (int(step["results/yx_touch"][1] * 1000), int(step["results/yx_touch"][0] * 1000))
        episode_data = {
            "episode_id": episode_data[0]["episode_id"],
            "goal_info": episode_data[0]["goal_info"],
            "steps": episode_data,
        }
        episode_data_list_new.append(episode_data)

    # window_size_list = [1, 2, 3]
    window_size_list = [0, 1, 2, 3]

    def process_episode(args):
        episode, window_size = args
        return process_android_episodes([episode], window_size)

    instructions = []
    for window_size in window_size_list:
        tasks = [(episode, window_size) for episode in episode_data_list_new]
        with multiprocessing.Pool(processes=multiprocessing.cpu_count()) as pool:
            results = list(tqdm(pool.imap(process_episode, tasks), total=len(tasks), desc=f"Window Size {window_size}"))
        for result in results:
            instructions.extend(result)

    print(f"Number of context aware train instructions: {len(instructions)}")
    
    with open(f"aitw_window_{'-'.join([str(e) for e in window_size_list])}_{len(instructions)//1000}k.json", "w", encoding="utf-8") as file:
        json.dump(instructions, file, ensure_ascii=False, indent=4)