stat.py

# + Group: data/demo_8
#   - Dataset: data/demo_8/actions, shape: (79, 7), dtype: float64
#   - Dataset: data/demo_8/dones, shape: (79,), dtype: uint8
#   + Group: data/demo_8/obs
#     - Dataset: data/demo_8/obs/agentview_rgb, shape: (79, 128, 128, 3), dtype: uint8
#     - Dataset: data/demo_8/obs/ee_ori, shape: (79, 3), dtype: float64
#     - Dataset: data/demo_8/obs/ee_pos, shape: (79, 3), dtype: float64
#     - Dataset: data/demo_8/obs/ee_states, shape: (79, 6), dtype: float64
#     - Dataset: data/demo_8/obs/eye_in_hand_rgb, shape: (79, 128, 128, 3), dtype: uint8
#     - Dataset: data/demo_8/obs/gripper_states, shape: (79, 2), dtype: float64
#     - Dataset: data/demo_8/obs/joint_states, shape: (79, 7), dtype: float64
#   - Dataset: data/demo_8/rewards, shape: (79,), dtype: uint8
#   - Dataset: data/demo_8/robot_states, shape: (79, 9), dtype: float64
#   - Dataset: data/demo_8/states, shape: (79, 92), dtype: float64
# + Group: data/demo_9
#   - Dataset: data/demo_9/actions, shape: (89, 7), dtype: float64
#   - Dataset: data/demo_9/dones, shape: (89,), dtype: uint8
#   + Group: data/demo_9/obs
#     - Dataset: data/demo_9/obs/agentview_rgb, shape: (89, 128, 128, 3), dtype: uint8
#     - Dataset: data/demo_9/obs/ee_ori, shape: (89, 3), dtype: float64
#     - Dataset: data/demo_9/obs/ee_pos, shape: (89, 3), dtype: float64
#     - Dataset: data/demo_9/obs/ee_states, shape: (89, 6), dtype: float64
#     - Dataset: data/demo_9/obs/eye_in_hand_rgb, shape: (89, 128, 128, 3), dtype: uint8
#     - Dataset: data/demo_9/obs/gripper_states, shape: (89, 2), dtype: float64
#     - Dataset: data/demo_9/obs/joint_states, shape: (89, 7), dtype: float64
#   - Dataset: data/demo_9/rewards, shape: (89,), dtype: uint8
#   - Dataset: data/demo_9/robot_states, shape: (89, 9), dtype: float64
#   - Dataset: data/demo_9/states, shape: (89, 92), dtype: float64

# The above is the structure of the HDF5 file. Read all the HDF5 files in the directory, and calculate the mean, std, min, max, q01, q99 of the actions, obs/ee_states, gripper_states, joint_states of all the files.

import h5py
import numpy as np
import os

def calculate_statistics(hdf5_path):
    actions = []
    ee_states = []
    gripper_states = []
    joint_states = []

    with h5py.File(hdf5_path, 'r') as f:
        for demo in f['data']:
            actions.append(f[f'data/{demo}/actions'][:])
            ee_states.append(f[f'data/{demo}/obs/ee_states'][:])
            gripper_states.append(f[f'data/{demo}/obs/gripper_states'][:])
            joint_states.append(f[f'data/{demo}/obs/joint_states'][:])
    actions = np.concatenate(actions, axis=0)
    ee_states = np.concatenate(ee_states, axis=0)
    gripper_states = np.concatenate(gripper_states, axis=0)
    joint_states = np.concatenate(joint_states, axis=0)
    stats = {
        'actions': {
            'mean': np.mean(actions, axis=0),
            'std': np.std(actions, axis=0),
            'min': np.min(actions, axis=0),
            'max': np.max(actions, axis=0),
            'q01': np.percentile(actions, 1, axis=0),
            'q99': np.percentile(actions, 99, axis=0)
        },
        'ee_states': {
            'mean': np.mean(ee_states, axis=0),
            'std': np.std(ee_states, axis=0),
            'min': np.min(ee_states, axis=0),
            'max': np.max(ee_states, axis=0),
            'q01': np.percentile(ee_states, 1, axis=0),
            'q99': np.percentile(ee_states, 99, axis=0)
        },
        'gripper_states': {
            'mean': np.mean(gripper_states, axis=0),
            'std': np.std(gripper_states, axis=0),
            'min': np.min(gripper_states, axis=0),
            'max': np.max(gripper_states, axis=0),
            'q01': np.percentile(gripper_states, 1, axis=0),
            'q99': np.percentile(gripper_states, 99, axis=0)
        },
        'joint_states': {
            'mean': np.mean(joint_states, axis=0),
            'std': np.std(joint_states, axis=0),
            'min': np.min(joint_states, axis=0),
            'max': np.max(joint_states, axis=0),
            'q01': np.percentile(joint_states, 1, axis=0),
            'q99': np.percentile(joint_states, 99, axis=0)
        }
    }
    return stats    

def process_directory(directory):
    all_stats = {
        'actions': [],
        'ee_states': [],
        'gripper_states': [],
        'joint_states': []
    }
    
    for filename in os.listdir(directory):
        if filename.endswith('.hdf5'):
            hdf5_path = os.path.join(directory, filename)
            stats = calculate_statistics(hdf5_path)
            for key in all_stats:
                all_stats[key].append(stats[key])
    
    # Calculate overall statistics
    overall_stats = {}
    for key, values in all_stats.items():
        # values: a list of dictionaries
        means = np.array([v['mean'] for v in values])
        stds = np.array([v['std'] for v in values])
        mins = np.array([v['min'] for v in values])
        maxs = np.array([v['max'] for v in values])
        q01s = np.array([v['q01'] for v in values])
        q99s = np.array([v['q99'] for v in values])
        overall_stats[key] = {
            'mean': np.mean(means, axis=0),
            'std': np.mean(stds, axis=0),
            'min': np.min(mins, axis=0),
            'max': np.max(maxs, axis=0),
            'q01': np.mean(q01s, axis=0),
            'q99': np.mean(q99s, axis=0)
        }

        
    return overall_stats

if __name__ == "__main__":
    directory = '/home2/czhang/datasets/LIBERO/libero_spatial'
    stats = process_directory(directory)
    for key, value in stats.items():
        print(f"{key}:")
        for stat_name, stat_value in value.items():
            print(f"  {stat_name}: {stat_value}")
    print("Statistics calculated successfully.")