import numpy as np import os import IPython from cliport import tasks from cliport.dataset import RavensDataset from cliport.environments.environment import Environment from pygments import highlight from pygments.lexers import PythonLexer from pygments.formatters import TerminalFormatter import time import random import json import traceback from gensim.utils import ( mkdir_if_missing, save_text, save_stat, compute_diversity_score_from_assets, add_to_txt, extract_dict, extract_code, extract_code_topdown ) import pybullet as p import copy dummy_task = {"task_name": "dummy", "task_descriptions": "dummy", "assets-used": "dummy"} class TopDownSimulationRunner: """ the main class that runs simulation loop """ def __init__(self, cfg, memory): self.cfg = cfg self.memory = memory # statistics self.syntax_pass_rate = 0 self.runtime_pass_rate = 0 self.env_pass_rate = 0 self.curr_trials = 0 self.prompt_folder = f"prompts/{cfg['prompt_folder']}" self.chat_log = memory.chat_log self.task_asset_logs = [] # All the generated tasks in this run. # Different from the ones in online buffer that can load from offline. self.generated_task_assets = [] self.generated_task_programs = [] self.generated_task_names = [] self.generated_tasks = [] self.passed_tasks = [] # accepted ones def print_current_stats(self): """ print the current statistics of the simulation design """ print("=========================================================") print(f"{self.cfg['prompt_folder']} Trial {self.curr_trials} SYNTAX_PASS_RATE: {(self.syntax_pass_rate / (self.curr_trials)) * 100:.1f}% RUNTIME_PASS_RATE: {(self.runtime_pass_rate / (self.curr_trials)) * 100:.1f}% ENV_PASS_RATE: {(self.env_pass_rate / (self.curr_trials)) * 100:.1f}%") print("=========================================================") def save_stats(self): """ save the final simulation statistics """ self.diversity_score = compute_diversity_score_from_assets(self.task_asset_logs, self.curr_trials) save_stat(self.cfg, self.cfg['model_output_dir'], self.generated_tasks, self.syntax_pass_rate / (self.curr_trials), self.runtime_pass_rate / (self.curr_trials), self.env_pass_rate / (self.curr_trials), self.diversity_score) print("Model Folder: ", self.cfg['model_output_dir']) print(f"Total {len(self.generated_tasks)} New Tasks:", [task['task-name'] for task in self.generated_tasks]) try: print(f"Added {len(self.passed_tasks)} Tasks:", self.passed_tasks) except: pass def task_creation(self, res): """ create the task through interactions of agent and critic """ self.task_creation_pass = True mkdir_if_missing(self.cfg['model_output_dir']) try: start_time = time.time() task_def = extract_dict(res, prefix="new_task") exec(task_def, globals()) self.generated_task = new_task except: to_print = highlight(f"{str(traceback.format_exc())}", PythonLexer(), TerminalFormatter()) print("Task Creation Exception:", to_print) self.generated_task = copy.deepcopy(dummy_task) try: self.generated_asset = {} self.generated_code, self.curr_task_name = extract_code_topdown(res) self.task_asset_logs.append(self.generated_task["assets-used"]) self.generated_task_name = self.generated_task["task-name"] self.generated_tasks.append(self.generated_task) self.generated_task_assets.append(self.generated_asset) self.generated_task_programs.append(self.generated_code) self.generated_task_names.append(self.generated_task_name) except: to_print = highlight(f"{str(traceback.format_exc())}", PythonLexer(), TerminalFormatter()) print("Task Code Creation Exception:", to_print) self.task_creation_pass = False # self.curr_task_name = self.generated_task['task-name'] print("task creation time {:.3f}".format(time.time() - start_time)) def setup_env(self): """ build the new task""" env = Environment( self.cfg['assets_root'], disp=self.cfg['disp'], shared_memory=self.cfg['shared_memory'], hz=480, record_cfg=self.cfg['record'] ) task = eval(self.curr_task_name)() task.mode = self.cfg['mode'] record = self.cfg['record']['save_video'] save_data = self.cfg['save_data'] # Initialize scripted oracle agent and dataset. expert = task.oracle(env) self.cfg['task'] = self.generated_task["task-name"] data_path = os.path.join(self.cfg['data_dir'], "{}-{}".format(self.generated_task["task-name"], task.mode)) dataset = RavensDataset(data_path, self.cfg, n_demos=0, augment=False) print(f"Saving to: {data_path}") print(f"Mode: {task.mode}") # Start video recording if record: env.start_rec(f'{dataset.n_episodes+1:06d}') return task, dataset, env, expert def run_one_episode(self, dataset, expert, env, task, episode, seed): """ run the new task for one episode """ add_to_txt( self.chat_log, f"================= TRIAL: {self.curr_trials}", with_print=True) record = self.cfg['record']['save_video'] np.random.seed(seed) random.seed(seed) print('Oracle demo: {}/{} | Seed: {}'.format(dataset.n_episodes + 1, self.cfg['n'], seed)) env.set_task(task) obs = env.reset() info = env.info reward = 0 total_reward = 0 # Rollout expert policy for _ in range(task.max_steps): act = expert.act(obs, info) episode.append((obs, act, reward, info)) lang_goal = info['lang_goal'] obs, reward, done, info = env.step(act) total_reward += reward print(f'Total Reward: {total_reward:.3f} | Done: {done} | Goal: {lang_goal}') if done: break episode.append((obs, None, reward, info)) return total_reward def simulate_task(self): """ simulate the created task and save demonstrations """ total_cnt = 0. reset_success_cnt = 0. env_success_cnt = 0. seed = 123 self.curr_trials += 1 if p.isConnected(): p.disconnect() if not self.task_creation_pass: print("task creation failure => count as syntax exceptions.") return # Check syntax and compilation-time error try: exec(self.generated_code, globals()) task, dataset, env, expert = self.setup_env() self.syntax_pass_rate += 1 except: to_print = highlight(f"{str(traceback.format_exc())}", PythonLexer(), TerminalFormatter()) save_text(self.cfg['model_output_dir'], self.generated_task_name + '_error', str(traceback.format_exc())) print("========================================================") print("Syntax Exception:", to_print) return try: # Collect environment and collect data from oracle demonstrations. while total_cnt <= self.cfg['max_env_run_cnt']: total_cnt += 1 episode = [] total_reward = self.run_one_episode(dataset, expert, env, task, episode, seed) reset_success_cnt += 1 env_success_cnt += total_reward > 0.99 self.runtime_pass_rate += 1 print("Runtime Test Pass!") # the task can actually be completed with oracle. 50% success rates are high enough. if env_success_cnt >= total_cnt / 2: self.env_pass_rate += 1 print("Environment Test Pass!") if self.cfg['save_memory']: self.memory.save_task_to_offline_topdown(self.generated_task, self.generated_code) print(f"added new task to offline: {self.generated_task['task-name']}") except: to_print = highlight(f"{str(traceback.format_exc())}", PythonLexer(), TerminalFormatter()) save_text(self.cfg['model_output_dir'], self.generated_task_name + '_error', str(traceback.format_exc())) print("========================================================") print("Runtime Exception:", to_print) self.memory.save_run(self.generated_task)