Datasets:

ethenj
/

MVISTA-4D

import pickle
import os
import h5py
import numpy as np
import cv2
import json
import torch
import os
import subprocess
import shutil

def rgbd_to_pointcloud(rgb, depth, intrinsic, extrinsic):
    height, width = depth.shape
    fx, fy = intrinsic[0,0], intrinsic[1,1]
    cx, cy = intrinsic[0,2], intrinsic[1,2]
    
    # Create a mesh grid of pixel coordinates
    u = np.arange(width)
    v = np.arange(height)
    u, v = np.meshgrid(u, v)
    
    # Flatten the arrays
    u = u.flatten()
    v = v.flatten()
    depth = depth.flatten()
    
    # Filter out points with zero depth
    valid = depth > 0
    u = u[valid]
    v = v[valid]
    depth = depth[valid]
    
    # Back-project to 3D camera coordinates
    x = (u - cx) * depth / fx
    y = (v - cy) * depth / fy
    z = depth
    
    points_camera = np.vstack((x, y, z)).T  # Shape: (N, 3)
    
    # Convert to homogeneous coordinates
    points_camera_homogeneous = np.hstack((points_camera, np.ones((points_camera.shape[0], 1))))  # Shape: (N, 4)
    
    # Transform to world coordinates
    points_world_homogeneous = points_camera_homogeneous @ np.linalg.inv(extrinsic).T  # Shape: (N, 4)
    
    points_world = points_world_homogeneous[:, :3] / points_world_homogeneous[:, 3][:, np.newaxis]  # Shape: (N, 3)
    
    colors = rgb[v, u, :]  # Shape: (N, 3)
    
    return points_world, colors

def load_cameras_from_hdf5(dataset_path):
    print(f"Loading dataset from {dataset_path}...")
    if not os.path.isfile(dataset_path):
        print(f"Dataset does not exist at \n{dataset_path}\n")
        # exit()
        raise FileNotFoundError(f"Dataset does not exist at \n{dataset_path}\n")

    with h5py.File(dataset_path, "r") as root:
        # def print_structure(name, obj):
        #     print(f"Name: {name}, Type: {'Group' if isinstance(obj, h5py.Group) else 'Dataset'}")
        # root.visititems(print_structure)
        # Name: observation/surround_camera_0, Type: Group
        # Name: observation/surround_camera_0/cam2world_gl, Type: Dataset
        # Name: observation/surround_camera_0/depth, Type: Dataset
        # Name: observation/surround_camera_0/extrinsic_cv, Type: Dataset
        # Name: observation/surround_camera_0/intrinsic_cv, Type: Dataset
        # Name: observation/surround_camera_0/rgb, Type: Dataset
        cam_dict = dict()
        # import time
        # time_start = time.time()
        for cam_name in root[f"/observation/"].keys():
            if "surround_camera_" in cam_name:
                cam_dict[cam_name] = dict()
                rgb_encoded = root[f"/observation/{cam_name}/rgb"][()]
                depth_encoded = root[f"/observation/{cam_name}/depth"][()] 
                rgb = None
                depth = None
                for i in range(rgb_encoded.shape[0]):
                    _rgb_encoded = rgb_encoded[i].rstrip(b"\0")
                    rgb_decoded = cv2.imdecode(np.frombuffer(buffer=_rgb_encoded, dtype=np.uint8), flags=cv2.IMREAD_COLOR)
                    # rgb_decoded = cv2.cvtColor(rgb_decoded, cv2.COLOR_BGR2RGB)
                    rgb_decoded = np.ascontiguousarray(rgb_decoded,dtype=np.float32) / 255.0

                    _depth_encoded = depth_encoded[i].rstrip(b"\0")
                    depth_decoded = cv2.imdecode(np.frombuffer(buffer=_depth_encoded, dtype=np.uint8), flags=cv2.IMREAD_UNCHANGED)
                    depth_decoded = np.ascontiguousarray(depth_decoded,dtype=np.float32) / 1000.0
                    
                    if rgb is None:
                        rgb = np.zeros((rgb_encoded.shape[0],) + rgb_decoded.shape, dtype=np.float32)
                        depth = np.zeros((depth_encoded.shape[0],) + depth_decoded.shape, dtype=np.float32)
                        
                    rgb[i] = rgb_decoded
                    depth[i] = depth_decoded
                
                cam_dict[cam_name]["rgb"] = rgb
                cam_dict[cam_name]["depth"] = depth
                
                cam_dict[cam_name]["intrinsic_cv"] = np.ascontiguousarray(root[f"/observation/{cam_name}/intrinsic_cv"][0],dtype=np.float64)
                cam_dict[cam_name]["extrinsic_cv"] = np.eye(4, dtype=np.float64)
                cam_dict[cam_name]["extrinsic_cv"][:3,:] = np.ascontiguousarray(root[f"/observation/{cam_name}/extrinsic_cv"][0],dtype=np.float64)
                cam_dict[cam_name]["cam2world_gl"] = np.ascontiguousarray(root[f"/observation/{cam_name}/cam2world_gl"][0],dtype=np.float64)
                
        # time_end = time.time()
        
        # print(f"Loaded {len(cam_dict)} cameras, time cost: {time_end - time_start:.2f}s")
    
    return cam_dict


# if __name__ == "__main__":
    
#     for cam_name in cam_dict:
#         print(f"Camera: {cam_name}")
#         print(f"{cam_dict[cam_name]['intrinsic_cv']}")
#         print(f"{cam_dict[cam_name]['extrinsic_cv']}")
#         print(f"{cam_dict[cam_name]['cam2world_gl']}")
#         print(f"RGB shape: {cam_dict[cam_name]['rgb'].shape}, Depth shape: {cam_dict[cam_name]['depth'].shape}")
        
#     import open3d as o3d
#     # do the pointcloud mapping
#     pcds = []
    
#     for cam_name in cam_dict:
#         rgb = cam_dict[cam_name]["rgb"]
#         depth = cam_dict[cam_name]["depth"]
#         intrinsic_cv = cam_dict[cam_name]["intrinsic_cv"]
#         extrinsic_cv = cam_dict[cam_name]["extrinsic_cv"]
#         cam2world_gl = cam_dict[cam_name]["cam2world_gl"]
#         points, colors = rgbd_to_pointcloud(rgb[0], depth[0], intrinsic_cv, extrinsic_cv)
#         pcds += [o3d.geometry.PointCloud(o3d.utility.Vector3dVector(points))]
#         pcds[-1].colors = o3d.utility.Vector3dVector(colors)
#         # break
        
#     # save the pointcloud to a ply file
#     pcd = pcds[0]
#     for i in range(1, len(pcds)):
#         pcd += pcds[i]
#     o3d.io.write_point_cloud("test.ply", pcd)

def find_data(base_path, max_episodes_per_task=None):
    data_list = []
    task_list = os.listdir(base_path)
    task_episode_count = {}
    for task_path in task_list:
        # print(f"Processing task: {task_path}")
        # print(os.path.join(base_path, task_path))
        task_episode_count.setdefault(task_path, 0)
        try: 
            embodiment_list = os.listdir(os.path.join(base_path, task_path))
            for embodiment_path in embodiment_list:
                try:        
                    config_list = os.listdir(os.path.join(base_path, task_path, embodiment_path))
                    for config_path in config_list:
                        # print(f"  Processing config: {config_path}")
                        scene_info_path = os.path.join(base_path, task_path, embodiment_path, config_path, "scene_info.json")
                        # print(f"    scene_info_path: {scene_info_path}")
                        if not os.path.exists(scene_info_path):
                            print(f"    scene_info_path {scene_info_path} does not exist, skip.")
                            continue
                        with open(scene_info_path, 'r') as f:
                            scene_info = json.load(f)
                        # print(scene_info)
                        for key, value in scene_info.items():
                            if (
                                max_episodes_per_task is not None
                                and task_episode_count[task_path] >= max_episodes_per_task
                            ):
                                continue
                            data_dict = dict()
                            data_dict["data_path"] = os.path.join(base_path, task_path, embodiment_path, config_path, "data")
                            data_dict["episode_name"] = key.replace("_", "")
                            data_dict["instructions_path"] = os.path.join(base_path, task_path, embodiment_path, config_path, "instructions",f"{data_dict['episode_name']}.json")
                            data_dict["metadata_path"] = os.path.join(data_dict["data_path"], f"{data_dict['episode_name']}.json")
                            data_dict["jointdata_path"] = os.path.join(data_dict["data_path"], f"{data_dict['episode_name']}_joint_data.npz")
                            data_dict["cache_path"] = os.path.join(base_path, task_path, embodiment_path, config_path, "cache", f"{data_dict['episode_name']}")
                            if not os.path.exists(data_dict["metadata_path"]):
                                print(f"    metadata_path {data_dict['metadata_path']} does not exist, skip.")
                                continue
                            data_list.append(data_dict)
                            task_episode_count[task_path] += 1
                        # for key in scene_info:
                        #     if "file_name" in scene_info[key]:
                        #         file_name = scene_info[key]["file_name"]
                        #         text = scene_info[key]["text"]
                except:
                    continue             
        except:
            continue
        
    print(f"Found {len(data_list)} data samples.")
    return data_list

def get_valid_candidates(k, N, diff, hard_diff=None):

    if hard_diff is not None:
        if hard_diff <= 0:
            raise ValueError("hard_diff must be a positive integer representing multiples of 30 degrees.")
        all_indices = np.arange(N)
        first = np.random.choice(all_indices)
        selected_list = []
        used = set()
        for offset in range(k):
            idx = (first + offset * hard_diff) % N
            if idx in used:
                raise ValueError(
                    f"hard_diff={hard_diff} does not provide {k} unique views for N={N}."
                )
            selected_list.append(idx)
            used.add(idx)
        selected_list.sort()
        return selected_list

    def circular_distance(a, b, N):
        return np.minimum(np.abs(a - b), N - np.abs(a - b))

    all_indices = np.arange(N)
    # print("All indices:", all_indices)
    selected_list = []

    # random select first index from all indices
    first = np.random.choice(all_indices)
    selected_list.append(first)

    for _ in range(k-1):
        available_indices = []
        for idx in all_indices:
            if idx in selected_list:
                continue
            distances = [circular_distance(idx, sel, N) for sel in selected_list]
            if not all(dist > diff for dist in distances):
                available_indices.append(idx)
    
        if len(available_indices) == 0:
            raise ValueError("No valid candidates available to select.")

        selected_list.append(np.random.choice(available_indices))

    selected_list.sort()

    far_view_idx = None
    for idx in all_indices:
        if idx in selected_list:
            continue
        distances = [circular_distance(idx, sel, N) for sel in selected_list]
        # print("Index:", idx, "Distances:", distances)

        min_2_distances = sorted(distances)[:2]
        min_dist = min_2_distances[0] + min_2_distances[1]

        if min_dist > diff:
            # available_indices.append(idx)
            far_view_idx = idx
            break
    # print("Selected list before rearranging:", selected_list)
    # print("Far view index:", far_view_idx)
    if far_view_idx is not None:
        start_idx = far_view_idx
        while True:
            if start_idx in selected_list:
                break
            start_idx = (start_idx + 1) % N
            # print("Adjusted start index:", start_idx)
        # rearramge the list to let start_idx be the first
        start_pos = selected_list.index(start_idx)
        selected_list = selected_list[start_pos:] + selected_list[:start_pos]

    return selected_list
        
# if __name__ == "__main__":
#     # task_list = os.listdir(base_path)
#     find_data(base_path)

def image_to_float_array(images, scale_factor=None):
    if not isinstance(images, np.ndarray):
        # images = np.array(images)
        images = np.array(images, dtype=np.float32)
    r = images[..., 0]
    g = images[..., 1]
    b = images[..., 2]
    # print (r.min(), r.max(), g.min(), g.max(), b.min(), b.max())
    # print(image_array.shape)
    # RGB image needs to be converted to 24 bit integer.
    # float_array = np.sum(image_array * [0, 256, 0], axis=-1)
    float_array = r * 65536.0 + g *256.0 + b * 1.0
    if scale_factor is None:
        scale_factor = 256000.0

    scaled_array = float_array / scale_factor
    return scaled_array

def get_instructions(instructions_file_path, num=3):
    instructions = []
    with open(instructions_file_path, 'r') as f:
        instructions_dict = json.load(f)
    num_instructions = len(instructions_dict["seen"])
    # ids = np.random.choice(num_instructions, num, replace=False)
    ids = range(num)
    # print(ids)
    # select the instructions according to the ids
    instructions = instructions_dict["seen"]
    instructions = [instructions[i] for i in ids]
    return instructions

def sample_frame_indices(total_frames: int, target_frames: int = 81) -> list:
    # if total_frames <= target_frames:
    #     raise ValueError("Total frames must be greater than the number of target frames (81).")
    
    # use linspace to generate evenly spaced indices
    indices = np.linspace(0, total_frames - 1, target_frames)
    
    # round and convert to integer indices
    indices = np.round(indices).astype(int)
    indices = np.clip(indices, 0, total_frames - 1)
    return indices.tolist()

def backproject_depth_to_cam_xyz(depth_t: torch.Tensor,
                             K: torch.Tensor) -> torch.Tensor:
    """
    depth_t: [B, H, W] 
    K:       [3, 3] 
    E:       [4, 4] 
    """
    device = depth_t.device
    dtype  = depth_t.dtype
    B, H, W = depth_t.shape

    if K.dim() == 2:  # [3,3] -> [B,3,3]
        K = K.to(device=device, dtype=dtype).unsqueeze(0).expand(B, -1, -1)
    else:
        K = K.to(device=device, dtype=dtype)
        assert K.shape[0] == B, "Batch K expects shape [B,3,3]"

    fx = K[:, 0, 0].view(B, 1, 1)
    fy = K[:, 1, 1].view(B, 1, 1)
    cx = K[:, 0, 2].view(B, 1, 1)
    cy = K[:, 1, 2].view(B, 1, 1)

    u = torch.arange(W, device=device, dtype=dtype).view(1, 1, W).expand(B, H, W)
    v = torch.arange(H, device=device, dtype=dtype).view(1, H, 1).expand(B, H, W)

    z = depth_t  # [B,H,W]
    valid = z > 0

    Xc = (u - cx) / fx * z
    Yc = (v - cy) / fy * z
    Zc = z

    Pc = torch.stack([Xc, Yc, Zc], dim=0) #[3, T, H, W]
    Pc = torch.where(valid.unsqueeze(0), Pc, torch.ones_like(Pc)*10) # filter out depth error

    # ones = torch.ones_like(Zc)
    # Pc = torch.stack([Xc, Yc, Zc, ones], dim=1).reshape(B, 4, -1)   # [B,3,N]
    # c2w = torch.linalg.inv(E)
    # Pw = torch.bmm(c2w, Pc)[:, :3]
    # Pw = Pw.view(B, 3, H, W)                                  # [B,3,H,W]
    # Pw = torch.where(valid.unsqueeze(1), Pw, torch.ones_like(Pw)*10000)

    return Pc

def normalize_xyz_to_NDC(xyz: torch.Tensor, ortho_bound, eps: float = 1e-6) -> torch.Tensor:
    """
      xyz:        [3, T, H, W]
    """
    l, r = ortho_bound['l'], ortho_bound['r']
    b, t = ortho_bound['b'], ortho_bound['t']
    n, f = ortho_bound['n'], ortho_bound['f']
    # rx, ry = ortho_bound['rx'], ortho_bound['ry']
    
    rx = max(r - l, eps)
    ry = max(t - b, eps)
    rz = max(abs(f - n), eps)

    x_ndc = (2.0 * xyz[0] - (r + l)) / (rx)
    y_ndc = (2.0 * xyz[1] - (t + b)) / (ry)
    z_ndc = (2.0 * xyz[2] - (f + n)) / ( (n - f) if (n - f) != 0 else 1e-8 )

    return torch.stack([x_ndc, y_ndc, z_ndc], dim=0)

# def inv_ortho_ndc_to_cam(ndc_xyz, l,r,b,t,n,f):
#     x = 0.5*(r-l)*ndc_xyz[...,0] + 0.5*(r+l)
#     y = 0.5*(t-b)*ndc_xyz[...,1] + 0.5*(t+b)
#     z = -0.5*(f-n)*ndc_xyz[...,2] + 0.5*(f+n)
#     return torch.stack([x,y,z], dim=-1)

def inv_ortho_ndc_to_cam(ndc_xyz, l,r,b,t,n,f, eps: float = 1e-6):
    rx = r - l
    ry = t - b
    rz = n - f

    x = 0.5 * rx * ndc_xyz[...,0] + 0.5 * (r + l)
    y = 0.5 * ry * ndc_xyz[...,1] + 0.5 * (t + b)
    z = 0.5 * rz * ndc_xyz[...,2] + 0.5 * (f + n)
    return torch.stack([x,y,z], dim=-1)

class Camera(object):
    def __init__(self, c2w=None, w2c=None):
        assert (c2w is not None) or (w2c is not None), "Either c2w or w2c must be provided."
        if c2w is None:
            c2w = np.linalg.inv(w2c)
        elif w2c is None:
            w2c = np.linalg.inv(c2w)
        self.c2w_mat = c2w
        self.w2c_mat = w2c

def images_to_video(imgs: np.ndarray, out_path: str, fps: float = 30.0, is_rgb: bool = True) -> None:
    if (not isinstance(imgs, np.ndarray) or imgs.ndim != 4 or imgs.shape[3] not in (3, 4)):
        raise ValueError("imgs must be a numpy.ndarray of shape (N, H, W, C), with C equal to 3 or 4.")
    os.makedirs(os.path.dirname(out_path), exist_ok=True)
    n_frames, H, W, C = imgs.shape
    if C == 3:
        pixel_format = "rgb24" if is_rgb else "bgr24"
    else:
        pixel_format = "rgba"
    ffmpeg = subprocess.Popen(
        [
            "ffmpeg",
            "-y",
            "-loglevel",
            "error",
            "-f",
            "rawvideo",
            "-pixel_format",
            pixel_format,
            "-video_size",
            f"{W}x{H}",
            "-framerate",
            str(fps),
            "-i",
            "-",
            "-pix_fmt",
            "yuv420p",
            "-vcodec",
            "libx264",
            "-crf",
            "23",
            f"{out_path}",
        ],
        stdin=subprocess.PIPE,
    )
    ffmpeg.stdin.write(imgs.tobytes())
    ffmpeg.stdin.close()
    if ffmpeg.wait() != 0:
        raise IOError(f"Cannot open ffmpeg. Please check the output path and ensure ffmpeg is supported.")

    # print(
    #     f"🎬 Video is saved to `{out_path}`, containing \033[94m{n_frames}\033[0m frames at {W}×{H} resolution and {fps} FPS."
    # )

def back_up_codes(src_dir, dst_dir, exclude_dirs=['__pycache__', '.git', 'data', 'logs', '.vscode']):
    os.makedirs(dst_dir, exist_ok=True)
    if exclude_dirs is None:
        exclude_dirs = []
    if not os.path.exists(dst_dir):
        os.makedirs(dst_dir)
    for item in os.listdir(src_dir):
        if item in exclude_dirs:
            continue
        s = os.path.join(src_dir, item)
        d = os.path.join(dst_dir, item)
        if os.path.isdir(s):
            back_up_codes(s, d, exclude_dirs)
        else:
            if s.endswith('.py') or s.endswith('.json') or s.endswith('.md'):
                shutil.copy2(s, d)


def find_rlbench_data(base_path):
    data_list = []
    task_list = os.listdir(base_path)
    scene_data_path = os.path.join(base_path, "scene_data.npz")
    if not os.path.exists(scene_data_path):
        print(f" rlbench: \t scene_data_path {scene_data_path} does not exist.")
        raise FileNotFoundError(f" rlbench: \t scene_data_path {scene_data_path} does not exist.")
    episode_num_list = {}
    for task_path in task_list:
        # print(f"Processing task: {task_path}")
        # print(os.path.join(base_path, task_path))
        try: 
            # embodiment_list = os.listdir(os.path.join(base_path, task_path))
            variant_list = os.listdir(os.path.join(base_path, task_path))
            episode_num_list[task_path] = 0
            for variant_path in variant_list:
            # for embodiment_path in embodiment_list:
                # try:        
                    variation_description_path = os.path.join(base_path, task_path, variant_path, "variation_descriptions.pkl")
                    # print(f" rlbench: \t processing variation_description_path: {variation_description_path}")
                    if not os.path.exists(variation_description_path):
                        print(f" rlbench: \t variation_description_path {variation_description_path} does not exist, skip.")
                        continue
                    else:
                        # print(f" rlbench: \t found variation_description_path: {variation_description_path}")
                        with open(variation_description_path, 'rb') as f:
                            variation_descriptions = pickle.load(f)
                            # print(f" rlbench: \t variation_descriptions for {variant_path}: {variation_descriptions}")
                    # config_list = os.listdir(os.path.join(base_path, task_path, embodiment_path))
                    episode_list: list[str] = os.listdir(os.path.join(base_path, task_path, variant_path, 'episodes'))
                    
                    for episode_path in episode_list:
                        # print(f"  Processing config: {config_path}")
                        # scene_info_path = os.path.join(base_path, task_path, embodiment_path, config_path, "scene_info.json")
                        data_path = os.path.join(base_path, task_path, variant_path, 'episodes', episode_path)
                        low_dim_path = os.path.join(data_path, 'low_dim_obs.npz')
                        # print(f" rlbench: \t processing low_dim_path: {low_dim_path}")
                        if not os.path.exists(low_dim_path):
                            print(f" rlbench: \t low_dim_path {low_dim_path} does not exist, skip.")
                            continue
                        # with open(low_dim_path, 'rb') as f:
                        #     print(f" rlbench: \t processing low_dim_path: {low_dim_path}")
                        #     low_dim_obs = pickle.load(f)
                        #     print(f" rlbench: \t low_dim_obs keys: {low_dim_obs.keys()}")
                        # with open(scene_info_path, 'r') as f:
                        #     scene_info = json.load(f)
                        # print(scene_info)
                        # for key, value in scene_info.items():
                        data_dict = dict()
                        data_dict["data_path"] = data_path
                        data_dict["episode_name"] = episode_path
                        data_dict["variation_name"] = variant_path
                        data_dict["instructions_path"] = variation_description_path
                        # data_dict["metadata_path"] = os.path.join(data_dict["data_path"], f"{data_dict['episode_name']}.json")
                        data_dict["jointdata_path"] = low_dim_path
                        data_dict["cache_path"] = os.path.join(base_path, task_path, variant_path, "cache", f"{data_dict['episode_name']}")
                        data_dict["task_name"] = task_path
                        data_dict["scene_data_path"] = scene_data_path
                        data_list.append(data_dict)
                        episode_num_list[task_path] += 1
                        # for key in scene_info:
                        #     if "file_name" in scene_info[key]:
                        #         file_name = scene_info[key]["file_name"]
                        #         text = scene_info[key]["text"]
                # except:
                #     continue             
        except:
            continue
        
    for task_name, episode_num in episode_num_list.items():
        print(f"  Task: {task_name}, Episode num: {episode_num}")
    print(f"Found {len(data_list)} data samples.")

    return data_list


def get_instructions_rlbench(instructions_file_path):
    with open(instructions_file_path, 'rb') as f:
        variation_descriptions = pickle.load(f)
    return variation_descriptions

	import pickle
	import os
	import h5py
	import numpy as np
	import cv2
	import json
	import torch
	import os
	import subprocess
	import shutil

	def rgbd_to_pointcloud(rgb, depth, intrinsic, extrinsic):
	height, width = depth.shape
	fx, fy = intrinsic[0,0], intrinsic[1,1]
	cx, cy = intrinsic[0,2], intrinsic[1,2]

	# Create a mesh grid of pixel coordinates
	u = np.arange(width)
	v = np.arange(height)
	u, v = np.meshgrid(u, v)

	# Flatten the arrays
	u = u.flatten()
	v = v.flatten()
	depth = depth.flatten()

	# Filter out points with zero depth
	valid = depth > 0
	u = u[valid]
	v = v[valid]
	depth = depth[valid]

	# Back-project to 3D camera coordinates
	x = (u - cx) * depth / fx
	y = (v - cy) * depth / fy
	z = depth

	points_camera = np.vstack((x, y, z)).T # Shape: (N, 3)

	# Convert to homogeneous coordinates
	points_camera_homogeneous = np.hstack((points_camera, np.ones((points_camera.shape[0], 1)))) # Shape: (N, 4)

	# Transform to world coordinates
	points_world_homogeneous = points_camera_homogeneous @ np.linalg.inv(extrinsic).T # Shape: (N, 4)

	points_world = points_world_homogeneous[:, :3] / points_world_homogeneous[:, 3][:, np.newaxis] # Shape: (N, 3)

	colors = rgb[v, u, :] # Shape: (N, 3)

	return points_world, colors

	def load_cameras_from_hdf5(dataset_path):
	print(f"Loading dataset from {dataset_path}...")
	if not os.path.isfile(dataset_path):
	print(f"Dataset does not exist at \n{dataset_path}\n")
	# exit()
	raise FileNotFoundError(f"Dataset does not exist at \n{dataset_path}\n")

	with h5py.File(dataset_path, "r") as root:
	# def print_structure(name, obj):
	# print(f"Name: {name}, Type: {'Group' if isinstance(obj, h5py.Group) else 'Dataset'}")
	# root.visititems(print_structure)
	# Name: observation/surround_camera_0, Type: Group
	# Name: observation/surround_camera_0/cam2world_gl, Type: Dataset
	# Name: observation/surround_camera_0/depth, Type: Dataset
	# Name: observation/surround_camera_0/extrinsic_cv, Type: Dataset
	# Name: observation/surround_camera_0/intrinsic_cv, Type: Dataset
	# Name: observation/surround_camera_0/rgb, Type: Dataset
	cam_dict = dict()
	# import time
	# time_start = time.time()
	for cam_name in root[f"/observation/"].keys():
	if "surround_camera_" in cam_name:
	cam_dict[cam_name] = dict()
	rgb_encoded = root[f"/observation/{cam_name}/rgb"][()]
	depth_encoded = root[f"/observation/{cam_name}/depth"][()]
	rgb = None
	depth = None
	for i in range(rgb_encoded.shape[0]):
	_rgb_encoded = rgb_encoded[i].rstrip(b"\0")
	rgb_decoded = cv2.imdecode(np.frombuffer(buffer=_rgb_encoded, dtype=np.uint8), flags=cv2.IMREAD_COLOR)
	# rgb_decoded = cv2.cvtColor(rgb_decoded, cv2.COLOR_BGR2RGB)
	rgb_decoded = np.ascontiguousarray(rgb_decoded,dtype=np.float32) / 255.0

	_depth_encoded = depth_encoded[i].rstrip(b"\0")
	depth_decoded = cv2.imdecode(np.frombuffer(buffer=_depth_encoded, dtype=np.uint8), flags=cv2.IMREAD_UNCHANGED)
	depth_decoded = np.ascontiguousarray(depth_decoded,dtype=np.float32) / 1000.0

	if rgb is None:
	rgb = np.zeros((rgb_encoded.shape[0],) + rgb_decoded.shape, dtype=np.float32)
	depth = np.zeros((depth_encoded.shape[0],) + depth_decoded.shape, dtype=np.float32)

	rgb[i] = rgb_decoded
	depth[i] = depth_decoded

	cam_dict[cam_name]["rgb"] = rgb
	cam_dict[cam_name]["depth"] = depth

	cam_dict[cam_name]["intrinsic_cv"] = np.ascontiguousarray(root[f"/observation/{cam_name}/intrinsic_cv"][0],dtype=np.float64)
	cam_dict[cam_name]["extrinsic_cv"] = np.eye(4, dtype=np.float64)
	cam_dict[cam_name]["extrinsic_cv"][:3,:] = np.ascontiguousarray(root[f"/observation/{cam_name}/extrinsic_cv"][0],dtype=np.float64)
	cam_dict[cam_name]["cam2world_gl"] = np.ascontiguousarray(root[f"/observation/{cam_name}/cam2world_gl"][0],dtype=np.float64)

	# time_end = time.time()

	# print(f"Loaded {len(cam_dict)} cameras, time cost: {time_end - time_start:.2f}s")

	return cam_dict


	# if __name__ == "__main__":

	# for cam_name in cam_dict:
	# print(f"Camera: {cam_name}")
	# print(f"{cam_dict[cam_name]['intrinsic_cv']}")
	# print(f"{cam_dict[cam_name]['extrinsic_cv']}")
	# print(f"{cam_dict[cam_name]['cam2world_gl']}")
	# print(f"RGB shape: {cam_dict[cam_name]['rgb'].shape}, Depth shape: {cam_dict[cam_name]['depth'].shape}")

	# import open3d as o3d
	# # do the pointcloud mapping
	# pcds = []

	# for cam_name in cam_dict:
	# rgb = cam_dict[cam_name]["rgb"]
	# depth = cam_dict[cam_name]["depth"]
	# intrinsic_cv = cam_dict[cam_name]["intrinsic_cv"]
	# extrinsic_cv = cam_dict[cam_name]["extrinsic_cv"]
	# cam2world_gl = cam_dict[cam_name]["cam2world_gl"]
	# points, colors = rgbd_to_pointcloud(rgb[0], depth[0], intrinsic_cv, extrinsic_cv)
	# pcds += [o3d.geometry.PointCloud(o3d.utility.Vector3dVector(points))]
	# pcds[-1].colors = o3d.utility.Vector3dVector(colors)
	# # break

	# # save the pointcloud to a ply file
	# pcd = pcds[0]
	# for i in range(1, len(pcds)):
	# pcd += pcds[i]
	# o3d.io.write_point_cloud("test.ply", pcd)

	def find_data(base_path, max_episodes_per_task=None):
	data_list = []
	task_list = os.listdir(base_path)
	task_episode_count = {}
	for task_path in task_list:
	# print(f"Processing task: {task_path}")
	# print(os.path.join(base_path, task_path))
	task_episode_count.setdefault(task_path, 0)
	try:
	embodiment_list = os.listdir(os.path.join(base_path, task_path))
	for embodiment_path in embodiment_list:
	try:
	config_list = os.listdir(os.path.join(base_path, task_path, embodiment_path))
	for config_path in config_list:
	# print(f" Processing config: {config_path}")
	scene_info_path = os.path.join(base_path, task_path, embodiment_path, config_path, "scene_info.json")
	# print(f" scene_info_path: {scene_info_path}")
	if not os.path.exists(scene_info_path):
	print(f" scene_info_path {scene_info_path} does not exist, skip.")
	continue
	with open(scene_info_path, 'r') as f:
	scene_info = json.load(f)
	# print(scene_info)
	for key, value in scene_info.items():
	if (
	max_episodes_per_task is not None
	and task_episode_count[task_path] >= max_episodes_per_task
	):
	continue
	data_dict = dict()
	data_dict["data_path"] = os.path.join(base_path, task_path, embodiment_path, config_path, "data")
	data_dict["episode_name"] = key.replace("_", "")
	data_dict["instructions_path"] = os.path.join(base_path, task_path, embodiment_path, config_path, "instructions",f"{data_dict['episode_name']}.json")
	data_dict["metadata_path"] = os.path.join(data_dict["data_path"], f"{data_dict['episode_name']}.json")
	data_dict["jointdata_path"] = os.path.join(data_dict["data_path"], f"{data_dict['episode_name']}_joint_data.npz")
	data_dict["cache_path"] = os.path.join(base_path, task_path, embodiment_path, config_path, "cache", f"{data_dict['episode_name']}")
	if not os.path.exists(data_dict["metadata_path"]):
	print(f" metadata_path {data_dict['metadata_path']} does not exist, skip.")
	continue
	data_list.append(data_dict)
	task_episode_count[task_path] += 1
	# for key in scene_info:
	# if "file_name" in scene_info[key]:
	# file_name = scene_info[key]["file_name"]
	# text = scene_info[key]["text"]
	except:
	continue
	except:
	continue

	print(f"Found {len(data_list)} data samples.")
	return data_list

	def get_valid_candidates(k, N, diff, hard_diff=None):

	if hard_diff is not None:
	if hard_diff <= 0:
	raise ValueError("hard_diff must be a positive integer representing multiples of 30 degrees.")
	all_indices = np.arange(N)
	first = np.random.choice(all_indices)
	selected_list = []
	used = set()
	for offset in range(k):
	idx = (first + offset * hard_diff) % N
	if idx in used:
	raise ValueError(
	f"hard_diff={hard_diff} does not provide {k} unique views for N={N}."
	)
	selected_list.append(idx)
	used.add(idx)
	selected_list.sort()
	return selected_list

	def circular_distance(a, b, N):
	return np.minimum(np.abs(a - b), N - np.abs(a - b))

	all_indices = np.arange(N)
	# print("All indices:", all_indices)
	selected_list = []

	# random select first index from all indices
	first = np.random.choice(all_indices)
	selected_list.append(first)

	for _ in range(k-1):
	available_indices = []
	for idx in all_indices:
	if idx in selected_list:
	continue
	distances = [circular_distance(idx, sel, N) for sel in selected_list]
	if not all(dist > diff for dist in distances):
	available_indices.append(idx)

	if len(available_indices) == 0:
	raise ValueError("No valid candidates available to select.")

	selected_list.append(np.random.choice(available_indices))

	selected_list.sort()

	far_view_idx = None
	for idx in all_indices:
	if idx in selected_list:
	continue
	distances = [circular_distance(idx, sel, N) for sel in selected_list]
	# print("Index:", idx, "Distances:", distances)

	min_2_distances = sorted(distances)[:2]
	min_dist = min_2_distances[0] + min_2_distances[1]

	if min_dist > diff:
	# available_indices.append(idx)
	far_view_idx = idx
	break
	# print("Selected list before rearranging:", selected_list)
	# print("Far view index:", far_view_idx)
	if far_view_idx is not None:
	start_idx = far_view_idx
	while True:
	if start_idx in selected_list:
	break
	start_idx = (start_idx + 1) % N
	# print("Adjusted start index:", start_idx)
	# rearramge the list to let start_idx be the first
	start_pos = selected_list.index(start_idx)
	selected_list = selected_list[start_pos:] + selected_list[:start_pos]

	return selected_list

	# if __name__ == "__main__":
	# # task_list = os.listdir(base_path)
	# find_data(base_path)

	def image_to_float_array(images, scale_factor=None):
	if not isinstance(images, np.ndarray):
	# images = np.array(images)
	images = np.array(images, dtype=np.float32)
	r = images[..., 0]
	g = images[..., 1]
	b = images[..., 2]
	# print (r.min(), r.max(), g.min(), g.max(), b.min(), b.max())
	# print(image_array.shape)
	# RGB image needs to be converted to 24 bit integer.
	# float_array = np.sum(image_array * [0, 256, 0], axis=-1)
	float_array = r * 65536.0 + g 256.0 + b 1.0
	if scale_factor is None:
	scale_factor = 256000.0

	scaled_array = float_array / scale_factor
	return scaled_array

	def get_instructions(instructions_file_path, num=3):
	instructions = []
	with open(instructions_file_path, 'r') as f:
	instructions_dict = json.load(f)
	num_instructions = len(instructions_dict["seen"])
	# ids = np.random.choice(num_instructions, num, replace=False)
	ids = range(num)
	# print(ids)
	# select the instructions according to the ids
	instructions = instructions_dict["seen"]
	instructions = [instructions[i] for i in ids]
	return instructions

	def sample_frame_indices(total_frames: int, target_frames: int = 81) -> list:
	# if total_frames <= target_frames:
	# raise ValueError("Total frames must be greater than the number of target frames (81).")

	# use linspace to generate evenly spaced indices
	indices = np.linspace(0, total_frames - 1, target_frames)

	# round and convert to integer indices
	indices = np.round(indices).astype(int)
	indices = np.clip(indices, 0, total_frames - 1)
	return indices.tolist()

	def backproject_depth_to_cam_xyz(depth_t: torch.Tensor,
	K: torch.Tensor) -> torch.Tensor:
	"""
	depth_t: [B, H, W]
	K: [3, 3]
	E: [4, 4]
	"""
	device = depth_t.device
	dtype = depth_t.dtype
	B, H, W = depth_t.shape

	if K.dim() == 2: # [3,3] -> [B,3,3]
	K = K.to(device=device, dtype=dtype).unsqueeze(0).expand(B, -1, -1)
	else:
	K = K.to(device=device, dtype=dtype)
	assert K.shape[0] == B, "Batch K expects shape [B,3,3]"

	fx = K[:, 0, 0].view(B, 1, 1)
	fy = K[:, 1, 1].view(B, 1, 1)
	cx = K[:, 0, 2].view(B, 1, 1)
	cy = K[:, 1, 2].view(B, 1, 1)

	u = torch.arange(W, device=device, dtype=dtype).view(1, 1, W).expand(B, H, W)
	v = torch.arange(H, device=device, dtype=dtype).view(1, H, 1).expand(B, H, W)

	z = depth_t # [B,H,W]
	valid = z > 0

	Xc = (u - cx) / fx * z
	Yc = (v - cy) / fy * z
	Zc = z

	Pc = torch.stack([Xc, Yc, Zc], dim=0) #[3, T, H, W]
	Pc = torch.where(valid.unsqueeze(0), Pc, torch.ones_like(Pc)*10) # filter out depth error

	# ones = torch.ones_like(Zc)
	# Pc = torch.stack([Xc, Yc, Zc, ones], dim=1).reshape(B, 4, -1) # [B,3,N]
	# c2w = torch.linalg.inv(E)
	# Pw = torch.bmm(c2w, Pc)[:, :3]
	# Pw = Pw.view(B, 3, H, W) # [B,3,H,W]
	# Pw = torch.where(valid.unsqueeze(1), Pw, torch.ones_like(Pw)*10000)

	return Pc

	def normalize_xyz_to_NDC(xyz: torch.Tensor, ortho_bound, eps: float = 1e-6) -> torch.Tensor:
	"""
	xyz: [3, T, H, W]
	"""
	l, r = ortho_bound['l'], ortho_bound['r']
	b, t = ortho_bound['b'], ortho_bound['t']
	n, f = ortho_bound['n'], ortho_bound['f']
	# rx, ry = ortho_bound['rx'], ortho_bound['ry']

	rx = max(r - l, eps)
	ry = max(t - b, eps)
	rz = max(abs(f - n), eps)

	x_ndc = (2.0 * xyz[0] - (r + l)) / (rx)
	y_ndc = (2.0 * xyz[1] - (t + b)) / (ry)
	z_ndc = (2.0 * xyz[2] - (f + n)) / ( (n - f) if (n - f) != 0 else 1e-8 )

	return torch.stack([x_ndc, y_ndc, z_ndc], dim=0)

	# def inv_ortho_ndc_to_cam(ndc_xyz, l,r,b,t,n,f):
	# x = 0.5(r-l)ndc_xyz[...,0] + 0.5*(r+l)
	# y = 0.5(t-b)ndc_xyz[...,1] + 0.5*(t+b)
	# z = -0.5(f-n)ndc_xyz[...,2] + 0.5*(f+n)
	# return torch.stack([x,y,z], dim=-1)

	def inv_ortho_ndc_to_cam(ndc_xyz, l,r,b,t,n,f, eps: float = 1e-6):
	rx = r - l
	ry = t - b
	rz = n - f

	x = 0.5 * rx * ndc_xyz[...,0] + 0.5 * (r + l)
	y = 0.5 * ry * ndc_xyz[...,1] + 0.5 * (t + b)
	z = 0.5 * rz * ndc_xyz[...,2] + 0.5 * (f + n)
	return torch.stack([x,y,z], dim=-1)

	class Camera(object):
	def __init__(self, c2w=None, w2c=None):
	assert (c2w is not None) or (w2c is not None), "Either c2w or w2c must be provided."
	if c2w is None:
	c2w = np.linalg.inv(w2c)
	elif w2c is None:
	w2c = np.linalg.inv(c2w)
	self.c2w_mat = c2w
	self.w2c_mat = w2c

	def images_to_video(imgs: np.ndarray, out_path: str, fps: float = 30.0, is_rgb: bool = True) -> None:
	if (not isinstance(imgs, np.ndarray) or imgs.ndim != 4 or imgs.shape[3] not in (3, 4)):
	raise ValueError("imgs must be a numpy.ndarray of shape (N, H, W, C), with C equal to 3 or 4.")
	os.makedirs(os.path.dirname(out_path), exist_ok=True)
	n_frames, H, W, C = imgs.shape
	if C == 3:
	pixel_format = "rgb24" if is_rgb else "bgr24"
	else:
	pixel_format = "rgba"
	ffmpeg = subprocess.Popen(
	[
	"ffmpeg",
	"-y",
	"-loglevel",
	"error",
	"-f",
	"rawvideo",
	"-pixel_format",
	pixel_format,
	"-video_size",
	f"{W}x{H}",
	"-framerate",
	str(fps),
	"-i",
	"-",
	"-pix_fmt",
	"yuv420p",
	"-vcodec",
	"libx264",
	"-crf",
	"23",
	f"{out_path}",
	],
	stdin=subprocess.PIPE,
	)
	ffmpeg.stdin.write(imgs.tobytes())
	ffmpeg.stdin.close()
	if ffmpeg.wait() != 0:
	raise IOError(f"Cannot open ffmpeg. Please check the output path and ensure ffmpeg is supported.")

	# print(
	# f"🎬 Video is saved to `{out_path}`, containing \033[94m{n_frames}\033[0m frames at {W}×{H} resolution and {fps} FPS."
	# )

	def back_up_codes(src_dir, dst_dir, exclude_dirs=['__pycache__', '.git', 'data', 'logs', '.vscode']):
	os.makedirs(dst_dir, exist_ok=True)
	if exclude_dirs is None:
	exclude_dirs = []
	if not os.path.exists(dst_dir):
	os.makedirs(dst_dir)
	for item in os.listdir(src_dir):
	if item in exclude_dirs:
	continue
	s = os.path.join(src_dir, item)
	d = os.path.join(dst_dir, item)
	if os.path.isdir(s):
	back_up_codes(s, d, exclude_dirs)
	else:
	if s.endswith('.py') or s.endswith('.json') or s.endswith('.md'):
	shutil.copy2(s, d)


	def find_rlbench_data(base_path):
	data_list = []
	task_list = os.listdir(base_path)
	scene_data_path = os.path.join(base_path, "scene_data.npz")
	if not os.path.exists(scene_data_path):
	print(f" rlbench: \t scene_data_path {scene_data_path} does not exist.")
	raise FileNotFoundError(f" rlbench: \t scene_data_path {scene_data_path} does not exist.")
	episode_num_list = {}
	for task_path in task_list:
	# print(f"Processing task: {task_path}")
	# print(os.path.join(base_path, task_path))
	try:
	# embodiment_list = os.listdir(os.path.join(base_path, task_path))
	variant_list = os.listdir(os.path.join(base_path, task_path))
	episode_num_list[task_path] = 0
	for variant_path in variant_list:
	# for embodiment_path in embodiment_list:
	# try:
	variation_description_path = os.path.join(base_path, task_path, variant_path, "variation_descriptions.pkl")
	# print(f" rlbench: \t processing variation_description_path: {variation_description_path}")
	if not os.path.exists(variation_description_path):
	print(f" rlbench: \t variation_description_path {variation_description_path} does not exist, skip.")
	continue
	else:
	# print(f" rlbench: \t found variation_description_path: {variation_description_path}")
	with open(variation_description_path, 'rb') as f:
	variation_descriptions = pickle.load(f)
	# print(f" rlbench: \t variation_descriptions for {variant_path}: {variation_descriptions}")
	# config_list = os.listdir(os.path.join(base_path, task_path, embodiment_path))
	episode_list: list[str] = os.listdir(os.path.join(base_path, task_path, variant_path, 'episodes'))

	for episode_path in episode_list:
	# print(f" Processing config: {config_path}")
	# scene_info_path = os.path.join(base_path, task_path, embodiment_path, config_path, "scene_info.json")
	data_path = os.path.join(base_path, task_path, variant_path, 'episodes', episode_path)
	low_dim_path = os.path.join(data_path, 'low_dim_obs.npz')
	# print(f" rlbench: \t processing low_dim_path: {low_dim_path}")
	if not os.path.exists(low_dim_path):
	print(f" rlbench: \t low_dim_path {low_dim_path} does not exist, skip.")
	continue
	# with open(low_dim_path, 'rb') as f:
	# print(f" rlbench: \t processing low_dim_path: {low_dim_path}")
	# low_dim_obs = pickle.load(f)
	# print(f" rlbench: \t low_dim_obs keys: {low_dim_obs.keys()}")
	# with open(scene_info_path, 'r') as f:
	# scene_info = json.load(f)
	# print(scene_info)
	# for key, value in scene_info.items():
	data_dict = dict()
	data_dict["data_path"] = data_path
	data_dict["episode_name"] = episode_path
	data_dict["variation_name"] = variant_path
	data_dict["instructions_path"] = variation_description_path
	# data_dict["metadata_path"] = os.path.join(data_dict["data_path"], f"{data_dict['episode_name']}.json")
	data_dict["jointdata_path"] = low_dim_path
	data_dict["cache_path"] = os.path.join(base_path, task_path, variant_path, "cache", f"{data_dict['episode_name']}")
	data_dict["task_name"] = task_path
	data_dict["scene_data_path"] = scene_data_path
	data_list.append(data_dict)
	episode_num_list[task_path] += 1
	# for key in scene_info:
	# if "file_name" in scene_info[key]:
	# file_name = scene_info[key]["file_name"]
	# text = scene_info[key]["text"]
	# except:
	# continue
	except:
	continue

	for task_name, episode_num in episode_num_list.items():
	print(f" Task: {task_name}, Episode num: {episode_num}")
	print(f"Found {len(data_list)} data samples.")

	return data_list


	def get_instructions_rlbench(instructions_file_path):
	with open(instructions_file_path, 'rb') as f:
	variation_descriptions = pickle.load(f)
	return variation_descriptions