create_dataset.py

import copy
import os, sys
import random
import uuid
import pickle
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
import d4rl
import gym
import numpy as np
import pyrallis
import torch
import torch.nn as nn
import torch.nn.functional as F
import wandb
from torch.distributions import Normal
from torch.optim.lr_scheduler import CosineAnnealingLR
from torch import autograd

import imageio
import yaml

TensorBatch = List[torch.Tensor]
DEFAULT_DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'

@dataclass
class TrainConfig:
    #############################
    ######### Experiment ########
    #############################
    percent_expert: float = 0.1
    env_1: str = "halfcheetah-random-v2"
    env_2: str = "halfcheetah-expert-v2"
    file_name: str = "halfcheetah-10%-expert-random.npy"

def get_trajectory_indices(terminals):
    # Finds the indices where each trajectory begins and ends
    end_indices = np.where(terminals)[0]
    if len(end_indices) == 0 or end_indices[-1] != len(terminals) - 1:
        # Ensure the last index is included as an end if not already
        end_indices = np.append(end_indices, len(terminals) - 1)
    start_indices = np.append(0, end_indices[:-1] + 1)
    return list(zip(start_indices, end_indices + 1))

def modify_dataset(dataset, expert_dataset):
    # Assuming terminals are stored as boolean where True indicates the end of a trajectory
    trajectory_indices = get_trajectory_indices(dataset['terminals'] + dataset['timeouts'])
    expert_trajectory_indices = get_trajectory_indices(expert_dataset['terminals'] + expert_dataset['timeouts'])

    # Determine how many trajectories to replace
    trajectories_to_replace = int(config.percent_expert * len(trajectory_indices))

    # Handle the case when there are no trajectories to replace
    if trajectories_to_replace == 0 or len(expert_trajectory_indices) == 0:
        print("No trajectories to replace or no expert trajectories available.")
        trajectories_to_replace = 0
        indices_replace_map = {}
    else:
        # Randomly choose trajectories to replace
        indices_to_replace = np.random.choice(len(trajectory_indices), trajectories_to_replace, replace=False)
        expert_indices = np.random.choice(len(expert_trajectory_indices), trajectories_to_replace, replace=False)

        # Create a mapping from dataset trajectory indices to expert trajectory indices
        indices_replace_map = dict(zip(indices_to_replace, expert_indices))

    mixed_dataset = {}
    keys = ['observations', 'actions', 'next_observations', 'rewards', 'terminals']

    # Replacement process
    for key in keys:
        mixed_data = []
        for i in range(len(trajectory_indices)):
            if i in indices_replace_map:
                # Use expert trajectory
                expert_traj_idx = indices_replace_map[i]
                expert_start, expert_end = expert_trajectory_indices[expert_traj_idx]
                data_to_append = expert_dataset[key][expert_start:expert_end]
            else:
                # Use original dataset trajectory
                start, end = trajectory_indices[i]
                data_to_append = dataset[key][start:end]
            mixed_data.append(data_to_append)
        if mixed_data:
            # Concatenate all data for the current key
            mixed_dataset[key] = np.concatenate(mixed_data, axis=0)
        else:
            # If mixed_data is empty, create an empty array with the appropriate shape
            mixed_dataset[key] = np.array([], dtype=dataset[key].dtype)

    # Combine 'init_states' from both datasets
    mixed_dataset['init_states'] = np.concatenate([dataset['init_states'], expert_dataset['init_states']], axis=0)

    return mixed_dataset

def modify_dataset_direct(dataset, expert_dataset):
    mixed_dataset = {}
    keys = ['observations', 'actions', 'next_observations', 'rewards', 'terminals']
    # replace config.percent_expert of the dataset with expert_dataset
    for key in keys:
        mixed_data = []
        for i in reversed(range(len(dataset[key]))):
            if i > len(expert_dataset[key]) - int(config.percent_expert * len(dataset[key])):
                data_to_append = expert_dataset[key][i - len(dataset[key]) + 1]
            else:
                data_to_append = dataset[key][i]
            mixed_data.append(data_to_append)
        if mixed_data:
            # Concatenate all data for the current key
            mixed_dataset[key] = np.array(mixed_data)
        else:
            # If mixed_data is empty, create an empty array with the appropriate shape
            mixed_dataset[key] = np.array([], dtype=dataset[key].dtype)

    # Combine 'init_states' from both datasets
    mixed_dataset['init_states'] = np.concatenate([dataset['init_states'], expert_dataset['init_states']], axis=0)

    mixed_dataset['transition_ids'] = np.arange(len(mixed_dataset["observations"]))

    return mixed_dataset


def qlearning_dataset(env, dataset=None, terminate_on_end=False, **kwargs):
    if dataset is None:
        dataset = env.get_dataset(**kwargs)
    
    init_obs_index = np.unique(np.concatenate((np.where(dataset['terminals'])[0][:-1] + 1, np.where(dataset['timeouts'])[0][:-1] + 1)))
    init_obs_ = dataset['observations'][init_obs_index]

    N = dataset['rewards'].shape[0]
    obs_ = []
    next_obs_ = []
    action_ = []
    reward_ = []
    done_ = []
    timeout_ = []
    task_horizon = []

    # The newer version of the dataset adds an explicit
    # timeouts field. Keep old method for backwards compatability.
    use_timeouts = False
    if 'timeouts' in dataset:
        use_timeouts = True

    episode_step = 0
    for i in range(N-1):
        obs = dataset['observations'][i].astype(np.float32)
        new_obs = dataset['observations'][i+1].astype(np.float32)
        action = dataset['actions'][i].astype(np.float32)
        reward = dataset['rewards'][i].astype(np.float32)
        done_bool = bool(dataset['terminals'][i])
        timeout_bool = bool(dataset['timeouts'][i])

        if use_timeouts:
            final_timestep = dataset['timeouts'][i]
        else:
            final_timestep = (episode_step == env._max_episode_steps - 1)
        if (not terminate_on_end) and final_timestep:
            # Skip this transition and don't apply terminals on the last step of an episode
            episode_step = 0
            continue
        if done_bool or final_timestep:
            episode_step = 0

        obs_.append(obs)
        next_obs_.append(new_obs)
        action_.append(action)
        reward_.append(reward)
        done_.append(done_bool)
        timeout_.append(timeout_bool)
        task_horizon.append(episode_step)
        episode_step += 1
    
    
    # add in return for each episode
    return_list = [0]
    length = [0]
    for i in range(len(done_)):
        return_list[-1] += reward_[i]
        length[-1] += 1
        if done_[i] or timeout_[i]:
            return_list.append(0)
            length.append(0)

    count = 0
    data_return_list = [0] * len(done_)
    for i in range(len(done_)):
        data_return_list[i] = return_list[count]
        if done_[i] or timeout_[i]:
            count +=1
    
    data_return_list = env.get_normalized_score(np.array(data_return_list)) * 100.0
    data_return_list = np.array(data_return_list)


    epi_obs = []
    epi_n_obs = []
    epi_terminals = []
    epi_rewards = []
    epi_returns = []
    epi_actions = []
    obs = []
    n_obs = []
    terminals = []
    rewards = []
    actions = []
    # task_horizon = []
    task_step = 0
    for i in range(len(done_)):
        obs.append(obs_[i])
        n_obs.append(next_obs_[i])
        terminals.append(done_[i])
        rewards.append(reward_[i])
        actions.append(action_[i])
        # task_horizon.append(task_step)
        task_step += 1
        if done_[i] or timeout_[i]:
            epi_obs.append(np.array(obs))
            epi_n_obs.append(np.array(n_obs))
            epi_terminals.append(np.array(terminals))
            epi_rewards.append(np.array(rewards))
            epi_returns.append(data_return_list[i])
            epi_actions.append(np.array(actions))
            obs = []
            n_obs = []
            terminals = []
            rewards = []
            actions = []
            task_step = 0

    transition_ids = np.arange(len(obs_))
    return {
        'observations': np.array(obs_),
        'actions': np.array(action_),
        'next_observations': np.array(next_obs_),
        'rewards': np.array(reward_),
        'terminals': np.array(done_),
        'timeouts': np.array(timeout_),
        'init_states': np.array(init_obs_),
        'transition_ids': transition_ids,
        'returns': data_return_list,
        'epi_obs': np.array(epi_obs, dtype=object),
        'epi_n_obs': np.array(epi_n_obs, dtype=object),
        'epi_terminals': np.array(epi_terminals, dtype=object),
        'epi_rewards': np.array(epi_rewards, dtype=object),
        'epi_returns': np.array(epi_returns, dtype=object),
        'epi_actions': np.array(epi_actions, dtype=object),
        'task_horizon':np.array(task_horizon, dtype=object),
    }


def create_dataset():
    env = gym.make(config.env_1)
    expert_env = gym.make(config.env_2)

    dataset = qlearning_dataset(env)
    expert_dataset = qlearning_dataset(expert_env)

    # mixed_dataset = modify_dataset(dataset, expert_dataset)
    mixed_dataset = modify_dataset_direct(dataset, expert_dataset)

    with open(config.file_name, 'wb') as f:
        pickle.dump(mixed_dataset, f)

if __name__ == "__main__":
    config = pyrallis.parse(config_class=TrainConfig)
    create_dataset()