data/maze/create_maze.py

import networkx as nx
import random
import os
import argparse
import numpy

def generate_maze(n, edge_prob):
    # Create a directed grid graph with random edge removal
    G = nx.DiGraph()
    for i in range(n*n):
        G.add_node(i)
    
    # Add edges with probability edge_prob
    for i in range(n):
        for j in range(n):
            node = i*n + j
            # # up
            # if i > 0 and random.random() < edge_prob:
            #     G.add_edge(node, (i-1)*n + j)
            #     G.add_edge((i-1)*n + j, node)
            # down
            if i < n-1 and random.random() < edge_prob:
                G.add_edge(node, (i+1)*n + j)
                G.add_edge((i+1)*n + j, node)
            # # left
            # if j > 0 and random.random() < edge_prob:
            #     G.add_edge(node, i*n + j-1)
            #     G.add_edge(i*n + j-1, node)
            # right
            if j < n-1 and random.random() < edge_prob:
                G.add_edge(node, i*n + j+1)
                G.add_edge(i*n + j+1, node)
    
    return G

def print_grid(G, n, file=None):
    def write_line(text):
        if file is None:
            print(text, end="")
        else:
            file.write(text)
    
    for i in range(n-1):
        # Print row edges
        for j in range(n):
            write_line("+")
            if j < n-1 and G.has_edge(i*n + j, i*n + j+1):
                write_line("---")
            elif j < n-1:
                write_line("   ")
        write_line("\n")
        
        # Print column edges
        for j in range(n):
            if G.has_edge(i*n + j, (i+1)*n + j):
                write_line("|")
            else:
                write_line(" ")
            if j < n-1:
                write_line("   ")
        write_line("\n")
       
    # Print bottom border
    for j in range(n):
        write_line("+")
        if j < n-1 and G.has_edge((n-1)*n + j, (n-1)*n + j+1):
            write_line("---")
        elif j < n-1:
            write_line("   ")
    write_line("\n")

def get_reachable_nodes(G, target_node):  
    # Get the transitive closure of the graph  
    TC = nx.transitive_closure(G)  
    # Find the predecessors in the transitive closure (nodes that can reach the target_node)  
    reachable_from = TC.predecessors(target_node)  
    return list(reachable_from)

def obtain_reachability():
    reachability = {}  
    pairs = 0
    for node in maze_graph.nodes():  
        reachability[node] = get_reachable_nodes(maze_graph, node)
        pairs += len(reachability[node])
    return reachability, pairs

def random_walk(source_node, target_node):
    stack = [source_node]
    visited = [] # to eliminate cycles

    while stack != []:
        cur_node = stack.pop()
        visited.append(cur_node)
        if cur_node == target_node:
            return visited
        
        adj = list(maze_graph.successors(cur_node))
        anc = list(reachability[target_node])
        anc.append(target_node)
        
        remaining = [element for element in adj if element in anc and element not in visited] #if we want the path to contain cycles, we should remove "and element not in visited"
    
        if len(remaining) == 0:
            return random_walk(source_node, target_node) # for non-DAGs

        next_node = random.choice(remaining)
        stack.append(next_node)

    return visited

def seq2act(path):
    actions = []
    for i in range(1, len(path)):
        diff = path[i] - path[i-1]
        if diff == -n:
            actions.append('N')
        elif diff == n:
            actions.append('S')
        elif diff == -1:
            actions.append('W')
        elif diff == 1:
            actions.append('E')
    return actions


def wall_directions(node):
    """Return the list of NESW directions that hit a wall from `node`.

    A direction is a "wall" when the adjacent in-grid cell exists but there is
    no edge to it in the maze graph (i.e. the move is illegal).
    """
    i, j = divmod(node, n)
    dirs = []
    # North
    if i > 0 and not maze_graph.has_edge(node, (i - 1) * n + j):
        dirs.append('N')
    # South
    if i < n - 1 and not maze_graph.has_edge(node, (i + 1) * n + j):
        dirs.append('S')
    # West
    if j > 0 and not maze_graph.has_edge(node, i * n + (j - 1)):
        dirs.append('W')
    # East
    if j < n - 1 and not maze_graph.has_edge(node, i * n + (j + 1)):
        dirs.append('E')
    return dirs


def corrupt_one_token(path):
    """Turn a correct node path into a wrong (wall-hitting) action sequence.

    Takes the correct actions of `path` and replaces ONE move (at a random
    position, not necessarily the last) with an illegal direction that hits a
    wall from that move's starting cell. The remaining tokens are kept as-is.
    Returns the corrupted action list, or None if no position can be corrupted.
    """
    actions = seq2act(path)
    if not actions:
        return None
    # actions[i] is the move taken from node path[i]; try positions in random
    # order until we find one that has a wall to bump into.
    idxs = list(range(len(actions)))
    random.shuffle(idxs)
    for i in idxs:
        walls = wall_directions(path[i])  # illegal dirs from path[i]; excludes the legal actions[i]
        if walls:
            d = random.choice(walls)
            return actions[:i] + [d] + actions[i + 1:]
    return None


def select_task(tasks_config, is_train=True):
    """
    Randomly select a task based on the configured percentages.
    For training: uses 'train' percentage from tasks_config
    For testing: uses 'test' percentage from tasks_config
    Returns the selected task ID.
    """
    key = 'train' if is_train else 'test'
    percentages = [(task_id, config[key]) for task_id, config in tasks_config.items()]
    
    # Generate random number between 0 and 100
    rand = random.random() * 100
    cumsum = 0
    
    for task_id, pct in percentages:
        cumsum += pct
        if rand < cumsum:
            return task_id
    
    # Fallback to the last task if rounding errors occur
    return percentages[-1][0]

def create_dataset(i, tasks_config):
    train_set = []
    test_set = []
    train_num_per_pair = max(i,1)
    for target_node in range(num_nodes):
        cnt = 0 # to avoid some target not appear in training dataset
        for source_node in range(num_nodes):
            if source_node == target_node:
                continue
            if (data[source_node][target_node] == 1):
                if maze_graph.has_edge(source_node, target_node):
                    task_id = select_task(tasks_config, is_train=True)
                    if task_id == 'A':
                        train_set.append(['A', source_node, target_node] + seq2act([source_node, target_node]))
                    else:
                        print(f"Error: Task {task_id} is not yet defined. Skipping training data generation for this entry.")
                        
                for ii in range(train_num_per_pair):
                    task_id = select_task(tasks_config, is_train=True)
                    if task_id == 'A':
                        train_set.append(['A', source_node, target_node] + seq2act(random_walk(source_node, target_node)) )
                    else:
                        print(f"Error: Task {task_id} is not yet defined. Skipping training data generation for this entry.")
                        
            if (data[source_node][target_node] == -1):
                task_id = select_task(tasks_config, is_train=False)
                if task_id == 'A':
                    test_set.append(['A', source_node, target_node] + seq2act(random_walk(source_node, target_node)))
                else:
                    print(f"Error: Task {task_id} is not yet defined. Skipping test data generation for this entry.")
                    
    return train_set, test_set

def add_x(train_set, test_set, tasks_config):
    cnt = 0
    for target_node in range(num_nodes):
        for source_node in range(num_nodes):
            if source_node == target_node:
                continue
            if source_node not in reachability[target_node]:
                cnt += 1

    prob_in_test = len(test_set) / cnt * 0.2
    prob_in_train = min(len(train_set) / cnt * 0.2, 1 - prob_in_test)
    train_repeat = max(int(len(train_set) / cnt * 0.15 / prob_in_train), 1)
    print(prob_in_train, prob_in_test, train_repeat)

    for target_node in range(num_nodes):
        for source_node in range(num_nodes):
            if source_node == target_node:
                continue
            if source_node not in reachability[target_node]:
                coin = random.random()
                if coin < prob_in_train:
                    for _ in range(train_repeat):
                        task_id = select_task(tasks_config, is_train=True)
                        if task_id == 'A':
                            train_set.append(['A', source_node, target_node, 'x'])
                        else:
                            print(f"Error: Task {task_id} is not yet defined. Skipping training data generation for this entry.")
                        
                elif coin > 1 - prob_in_test:
                    task_id = select_task(tasks_config, is_train=False)
                    if task_id == 'A':
                        test_set.append(['A', source_node, target_node, 'x'])
                    else:
                        print(f"Error: Task {task_id} is not yet defined. Skipping test data generation for this entry.")

    return train_set, test_set


def add_wrong_paths(train_set, test_set, tasks_config, wrong_ratio):
    """Append wall-hitting wrong paths to the datasets.

    Each wrong path is built from a CORRECT path (source -> target) by replacing
    ONE move token (at a random position, not necessarily the last) with an
    illegal direction that hits a wall, then appending 'x' at the end.
    Format: ['A', source, target, <moves with one illegal move somewhere>, 'x'].
    During training only the final 'x' token is supervised (see train_maze.py /
    get_batch); every other token is masked out of the loss, so the model is
    not taught to imitate the wrong trajectory and must inspect the whole
    sequence to flag that a wall was hit.
    """
    if wrong_ratio <= 0:
        return train_set, test_set

    num_train_wrong = int(len(train_set) * wrong_ratio)
    num_test_wrong = int(len(test_set) * wrong_ratio)

    # Collect reachable (source, target) pairs, split like the correct data.
    train_pairs = []
    test_pairs = []
    for target_node in range(num_nodes):
        for source_node in range(num_nodes):
            if source_node == target_node:
                continue
            if source_node in reachability[target_node]:
                if data[source_node][target_node] == 1:
                    train_pairs.append((source_node, target_node))
                elif data[source_node][target_node] == -1:
                    test_pairs.append((source_node, target_node))

    def gen_wrong(pairs, count, is_train):
        out = []
        if not pairs or count <= 0:
            return out
        attempts = 0
        max_attempts = count * 50 + 100
        while len(out) < count and attempts < max_attempts:
            attempts += 1
            source_node, target_node = random.choice(pairs)
            path = random_walk(source_node, target_node)
            if not path or len(path) < 2:
                continue
            corrupted = corrupt_one_token(path)
            if corrupted is None:
                continue
            task_id = select_task(tasks_config, is_train=is_train)
            if task_id != 'A':
                continue
            out.append(['A', source_node, target_node] + corrupted + ['x'])
        return out

    train_wrong = gen_wrong(train_pairs, num_train_wrong, True)
    test_wrong = gen_wrong(test_pairs, num_test_wrong, False)
    print(f'Added {len(train_wrong)} wrong (wall-hit) paths to train, {len(test_wrong)} to test.')
    return train_set + train_wrong, test_set + test_wrong


def obtain_stats(dataset):
    max_len = 0
    pairs = set()

    for data in dataset:
        max_len = max(max_len, len(data))
        pairs.add((data[0],data[-1]))

    len_stats = [0]*(max_len + 1) 
    
    for data in dataset:
        length = len(data)
        len_stats[length] += 1
        
    print('number of source target pairs:', len(pairs))
    for ii in range(3, len(len_stats)):
        print(f'There are {len_stats[ii]} paths with length {ii-3}')


def format_data(data):
    # Format: task_id source target [remaining_tokens]
    return ' '.join(str(token) for token in data) + '\n'
        
def write_dataset(dataset, file_name):
    with open(file_name, "w") as file:
        for data in dataset:
            file.write(format_data(data))


def parse_tasks(tasks_str):
    """
    Parse task specification string into a dictionary.
    Format: "TaskID:train_percent:test_percent,TaskID:train_percent:test_percent,..."
    Example: "A:50:50,B:30:30,C:20:20"
    Returns: {"A": {"train": 50, "test": 50}, "B": {"train": 30, "test": 30}, ...}
    Validates that all training percentages sum to 100% and all test percentages sum to 100%.
    """
    tasks = {}
    for task_spec in tasks_str.split(','):
        parts = task_spec.strip().split(':')
        if len(parts) != 3:
            raise ValueError(f"Invalid task specification: {task_spec}. Expected format: TaskID:train_percent:test_percent")
        task_id, train_pct, test_pct = parts[0].strip(), int(parts[1].strip()), int(parts[2].strip())
        if task_id in tasks:
            raise ValueError(f"Duplicate task ID: {task_id}")
        tasks[task_id] = {"train": train_pct, "test": test_pct}
    
    # Validate that percentages sum to 100%
    total_train = sum(config["train"] for config in tasks.values())
    total_test = sum(config["test"] for config in tasks.values())
    
    if total_train != 100:
        raise ValueError(f"Training task percentages must sum to 100%, but got {total_train}%")
    if total_test != 100:
        raise ValueError(f"Test task percentages must sum to 100%, but got {total_test}%")
    
    return tasks


if __name__ == "__main__":  
    parser = argparse.ArgumentParser(description='Generate a maze based on the given parameters.')  
    parser.add_argument('--grid_size', type=int, default=10, help='Size of the grid (n x n)')  
    parser.add_argument('--edge_prob', type=float, default=0.6, help='Probability to keep an edge in the grid graph')  
    parser.add_argument('--chance_in_train', type=float, default=0.5, help='Chance of a pair being in the training set')  
    parser.add_argument('--num_of_paths', type=int, default=20, help='Number of paths per pair nodes in training dataset')  
    parser.add_argument('--wrong_ratio', type=float, default=0.0,
                        help='Fraction of extra wall-hitting wrong paths to add, relative to the '
                             'number of correct paths (e.g. 0.2 = +20%%). Each wrong path ends in '
                             'an illegal move + "x"; only the "x" step is supervised during training. '
                             'Default 0.0 (disabled).')  
    # Multi-task specification: comma-separated task identifiers with their train/test percentages
    # Format: "TaskID:train_percent:test_percent,TaskID:train_percent:test_percent,..."
    # Example: "A:100:100" or "A:50:50,B:30:30,C:20:20"
    # Default is Task A (pathfinding) with 100% in both train and test datasets
    parser.add_argument('--tasks', type=str, default='A:100:100', 
                        help='Task identifiers with percentages. Format: "TaskID:train_pct:test_pct,TaskID:train_pct:test_pct,..." (default: A:100:100)')  

    args = parser.parse_args()  
    
    # Parse task specifications
    tasks_config = parse_tasks(args.tasks)

    n = args.grid_size
    edge_prob = args.edge_prob
    num_nodes = n * n
    chance_in_train = args.chance_in_train
    num_of_paths = args.num_of_paths

    maze_graph = generate_maze(n, edge_prob)
    
    folder_name = os.path.join(os.path.dirname(__file__), f'{num_nodes}')
    if not os.path.exists(folder_name):  
        os.makedirs(folder_name)
    
    # Save grid visualization to file
    grid_file_path = os.path.join(folder_name, f'maze_{n}_{edge_prob}_{num_of_paths}.txt')
    with open(grid_file_path, 'w') as f:
        print_grid(maze_graph, n, file=f)
    
    print_grid(maze_graph, n)
    reachability, feasible_pairs = obtain_reachability()

    data = numpy.zeros([num_nodes,num_nodes])
    for target_node in range(num_nodes):
        cnt = 0 # to avoid some target not appear in training dataset
        for source_node in range(num_nodes):
            if source_node == target_node:
                continue
            if source_node in reachability[target_node]:
                if (maze_graph.has_edge(source_node, target_node)) or random.random() < chance_in_train or cnt < 1:
                    data[source_node][target_node] = 1
                    cnt += 1 
                else:
                    data[source_node][target_node] = -1
                
    train_set, test_set = create_dataset(num_of_paths, tasks_config)

    train_set, test_set = add_wrong_paths(train_set, test_set, tasks_config, args.wrong_ratio)

    obtain_stats(train_set)
    print('number of source target pairs:', len(test_set))

    write_dataset(train_set, os.path.join(folder_name, f'train_{num_of_paths}.txt'))
    write_dataset(test_set, os.path.join(folder_name, f'test.txt'))
    nx.write_graphml(maze_graph, os.path.join(folder_name, f'maze_graph.graphml'))