def main(): ap = argparse.ArgumentParser() ap.add_argument('--arch', required=True, nargs='+', type=int) ap.add_argument('--lr', required=False, type=int, default=.01) ap.add_argument('--epochs', required=False, type=int, default=100) ap.add_argument('--iters', required=True, type=int) ap.add_argument('--trials', required=True, type=int) ap.add_argument('--env', required=True) ap.add_argument('--t', required=True, type=int) ap.add_argument('--grads', required=True, type=int) ap.add_argument('--weights', required=True, nargs='+', type=float, default=[1.0, .1, .5]) ap.add_argument('--ufact', required=True, default=4.0, type=float) ap.add_argument('--id', required=True, default=4.0, type=int) opt = Options() args = ap.parse_args() opt.load_args(args) args = vars(args) opt.envname = opt.env opt.env = gym.envs.make(opt.envname).env opt.sim = gym.envs.make(opt.envname).env exp_id = args['id'] opt.env.my_weights = args['weights'] opt.env.ufact = args['ufact'] opt.pi = net.Network([64, 64], .01, 300) suffix = '_' + utils.stringify(args['weights']) + '_' + str(args['ufact']) weights_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str( exp_id) + '_weights' + suffix + '.txt' stats_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str( exp_id) + '_stats' + suffix + '.txt' opt.pi.load_weights(weights_path, stats_path) opt.sup = Supervisor(opt.pi) opt.misc = Options() opt.misc.num_evaluations = 1 try: for t in range(opt.trials): start_time = timer.time() results = run_trial(opt) except KeyboardInterrupt: pass
def main(): ap = argparse.ArgumentParser() ap.add_argument('--arch', required=True, nargs='+', type=int) ap.add_argument('--lr', required=False, type=float, default=.01) ap.add_argument('--epochs', required=False, type=int, default=100) ap.add_argument('--iters', required=True, type=int) ap.add_argument('--trials', required=True, type=int) ap.add_argument('--env', required=True) ap.add_argument('--t', required=True, type=int) ap.add_argument('--grads', required=True, type=int) ap.add_argument('--weights', required=True, nargs='+', type=float, default=[1.0, .1, .5]) ap.add_argument('--ufact', required=True, default=4.0, type=float) ap.add_argument('--id', required=True, default=4.0, type=int) ap.add_argument('--nu', required=True, type=float) ap.add_argument('--gamma', required=True, type=float) opt = Options() args = ap.parse_args() opt.load_args(args) args = vars(args) opt.envname = opt.env opt.env = gym.envs.make(opt.envname).env opt.sim = gym.envs.make(opt.envname).env exp_id = args['id'] opt.env.my_weights = args['weights'] opt.env.ufact = args['ufact'] opt.pi = net.Network([64, 64], .01, 300) suffix = '_' + utils.stringify(args['weights']) + '_' + str(args['ufact']) weights_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str( exp_id) + '_weights' + suffix + '.txt' stats_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str( exp_id) + '_stats' + suffix + '.txt' opt.pi.load_weights(weights_path, stats_path) opt.sup = Supervisor(opt.pi) opt.misc = Options() opt.misc.num_evaluations = 1 opt.misc.samples = 5 rec_results = {} lnr_results = {} plot_dir = utils.generate_plot_dir('initial', 'experts', vars(opt)) data_dir = utils.generate_data_dir('initial', 'experts', vars(opt)) if not os.path.exists(plot_dir): os.makedirs(plot_dir) if not os.path.exists(data_dir): os.makedirs(data_dir) if not os.path.exists(data_dir + '/full'): os.makedirs(data_dir + '/full') if not os.path.exists(plot_dir + '/full'): os.makedirs(plot_dir + '/full') opt.data_dir = data_dir opt.plot_dir = plot_dir trials_data = run_trial(opt)
def main(): ap = argparse.ArgumentParser() ap.add_argument('--arch', required=True, nargs='+', type=int) ap.add_argument('--lr', required=False, type=float, default=.01) ap.add_argument('--epochs', required=False, type=int, default=100) ap.add_argument('--iters', required=True, type=int) ap.add_argument('--trials', required=True, type=int) ap.add_argument('--env', required=True) ap.add_argument('--t', required=True, type=int) ap.add_argument('--grads', required=True, type=int) ap.add_argument('--weights', required=True, nargs='+', type=float, default=[1.0, .1, .5]) ap.add_argument('--ufact', required=True, default=4.0, type=float) ap.add_argument('--id', required=True, default=4.0, type=int) ap.add_argument('--nu', required=True, type=float) ap.add_argument('--gamma', required=True, type=float) opt = Options() args = ap.parse_args() opt.load_args(args) args = vars(args) opt.envname = opt.env opt.env = gym.envs.make(opt.envname).env opt.sim = gym.envs.make(opt.envname).env exp_id = args['id'] opt.env.my_weights = args['weights'] opt.env.ufact = args['ufact'] opt.pi = net.Network([64, 64], .01, 300) suffix = '_' + utils.stringify(args['weights']) + '_' + str(args['ufact']) weights_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str(exp_id) + '_weights' + suffix + '.txt' stats_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str(exp_id) + '_stats' + suffix + '.txt' opt.pi.load_weights(weights_path, stats_path) opt.sup = Supervisor(opt.pi) opt.misc = Options() opt.misc.num_evaluations = 1 opt.misc.samples = 5 rec_results = {} lnr_results = {} plot_dir = utils.generate_plot_dir('initial', 'experts', vars(opt)) data_dir = utils.generate_data_dir('initial', 'experts', vars(opt)) if not os.path.exists(plot_dir): os.makedirs(plot_dir) if not os.path.exists(data_dir): os.makedirs(data_dir) if not os.path.exists(data_dir + '/opt'): os.makedirs(data_dir + '/opt') if not os.path.exists(plot_dir + '/opt'): os.makedirs(plot_dir + '/opt') trials_data = [] try: for t in range(opt.trials): print "Trial: " + str(t) start_time = timer.time() results = run_trial(opt) trials_data.append(results) pickle.dump(trials_data, open(data_dir + 'opt/trials_data.pkl', 'w')) except KeyboardInterrupt: pass
envname = 'Pusher6-v0' exp_id = args['id'] env = gym.envs.make(envname).env env.my_weights = args['weights'] env.ufact = args['ufact'] net = net.Network([64, 64], .01, 300) #net.load_weights('meta/weights.txt', 'meta/stats.txt') suffix = '_' + utils.stringify(args['weights']) + '_' + str(args['ufact']) weights_path = 'meta/' + 'test' + '/' + envname + '_' + str( exp_id) + '_weights' + suffix + '.txt' stats_path = 'meta/' + 'test' + '/' + envname + '_' + str( exp_id) + '_stats' + suffix + '.txt' net.load_weights(weights_path, stats_path) net_sup = Supervisor(net) opt = Options opt.env = env opt.sup = net_sup opt.t = 100 est = knet.Network([64, 64], learning_rate=.01, epochs=100) lnr = learner.Learner(est) oc = OneClassSVM(kernel='rbf', gamma=.01, nu=.01) ITERATIONS = 500 print "\n\nSup rollouts\n\n" sup_failures = 0 initial_states = []
def main(): ap = argparse.ArgumentParser() ap.add_argument('--arch', required=True, nargs='+', type=int) ap.add_argument('--lr', required=False, type=float, default=.01) ap.add_argument('--epochs', required=False, type=int, default=100) ap.add_argument('--iters', required=True, type=int) ap.add_argument('--trials', required=True, type=int) ap.add_argument('--env', required=True) ap.add_argument('--t', required=True, type=int) ap.add_argument('--grads', required=True, type=int) ap.add_argument('--weights', required=True, nargs='+', type=float, default=[1.0, .1, .5]) ap.add_argument('--ufact', required=True, default=4.0, type=float) ap.add_argument('--id', required=True, default=4.0, type=int) ap.add_argument('--nu', required=True, type=float) ap.add_argument('--gamma', required=True, type=float) opt = Options() args = ap.parse_args() opt.load_args(args) args = vars(args) opt.envname = opt.env opt.env = gym.envs.make(opt.envname).env opt.sim = gym.envs.make(opt.envname).env exp_id = args['id'] opt.env.my_weights = args['weights'] opt.env.ufact = args['ufact'] opt.pi = net.Network([64, 64], .01, opt.epochs) suffix = '_' + utils.stringify(args['weights']) + '_' + str(args['ufact']) weights_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str(exp_id) + '_weights' + suffix + '.txt' stats_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str(exp_id) + '_stats' + suffix + '.txt' opt.pi.load_weights(weights_path, stats_path) opt.sup = Supervisor(opt.pi) opt.misc = Options() opt.misc.num_evaluations = 1 opt.misc.samples = 1 rec_results = {} lnr_results = {} plot_dir = utils.generate_plot_dir('initial', 'experts', vars(opt)) data_dir = utils.generate_data_dir('initial', 'experts', vars(opt)) if not os.path.exists(plot_dir): os.makedirs(plot_dir) if not os.path.exists(data_dir): os.makedirs(data_dir) if not os.path.exists(data_dir + '/multiple_trials'): os.makedirs(data_dir + '/multiple_trials') if not os.path.exists(plot_dir + '/multiple_trials'): os.makedirs(plot_dir + '/multiple_trials') opt.data_dir = data_dir opt.plot_dir = plot_dir trials_data = [] rec_scores = [] rec_cutoffs = [] for t in range(opt.trials): print "\n\nTrial: " + str(t) + "\n\n" opt.t_value = t trial_data, info = run_trial(opt) trials_data.append(trial_data) rec_scores += info['rec_scores'] rec_cutoffs += info['rec_cutoffs'] print "Saving to: " + str(opt.data_dir + 'multiple_trials/trials_data.pkl') pickle.dump(trials_data, open(opt.data_dir + 'multiple_trials/trials_data.pkl', 'w')) pickle.dump(rec_scores, open(opt.data_dir + 'multiple_trials/rec_scores.pkl', 'w')) pickle.dump(rec_cutoffs, open(opt.data_dir + 'multiple_trials/rec_cutoffs.pkl', 'w'))
def main(): ap = argparse.ArgumentParser() ap.add_argument('--arch', required=True, nargs='+', type=int) ap.add_argument('--lr', required=False, type=int, default=.01) ap.add_argument('--epochs', required=False, type=int, default=100) ap.add_argument('--iters', required=True, type=int) ap.add_argument('--trials', required=True, type=int) ap.add_argument('--env', required=True) ap.add_argument('--t', required=True, type=int) ap.add_argument('--grads', required=True, type=int) ap.add_argument('--weights', required=True, nargs='+', type=float, default=[1.0, .1, .5]) ap.add_argument('--ufact', required=True, default=4.0, type=float) ap.add_argument('--id', required=True, default=4.0, type=int) opt = Options() args = ap.parse_args() opt.load_args(args) args = vars(args) opt.envname = opt.env opt.env = gym.envs.make(opt.envname).env opt.sim = gym.envs.make(opt.envname).env exp_id = args['id'] opt.env.my_weights = args['weights'] opt.env.ufact = args['ufact'] opt.pi = net.Network([64, 64], .01, 300) suffix = '_' + utils.stringify(args['weights']) + '_' + str(args['ufact']) weights_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str( exp_id) + '_weights' + suffix + '.txt' stats_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str( exp_id) + '_stats' + suffix + '.txt' opt.pi.load_weights(weights_path, stats_path) opt.sup = Supervisor(opt.pi) opt.misc = Options() opt.misc.num_evaluations = 1 opt.misc.samples = 600 rec_results = {} lnr_results = {} try: for t in range(opt.trials): start_time = timer.time() results = run_trial(opt) for key in results['rec'].keys(): if key in rec_results: rec_results[key].append(results['rec'][key]) else: rec_results[key] = [results['rec'][key]] for key in results['lnr'].keys(): if key in lnr_results: lnr_results[key].append(results['lnr'][key]) else: lnr_results[key] = [results['lnr'][key]] except KeyboardInterrupt: pass labels = sorted(list(rec_results.keys())) for key in rec_results.keys(): rec_results[key] = np.array(rec_results[key]) / float(opt.misc.samples) lnr_results[key] = np.array(lnr_results[key]) / float(opt.misc.samples) rec_means = [ np.mean(rec_results[key]) for key in sorted(rec_results.keys()) ] rec_sems = [ scipy.stats.sem(rec_results[key]) for key in sorted(rec_results.keys()) ] lnr_means = [ np.mean(lnr_results[key]) for key in sorted(lnr_results.keys()) ] lnr_sems = [ scipy.stats.sem(lnr_results[key]) for key in sorted(lnr_results.keys()) ] plt.style.use('ggplot') width = .4 index = np.arange(len(rec_means)) plt.bar(index + 0 * width, rec_means, width, label='recovery') plt.bar(index + 1 * width, lnr_means, width, label='no recovery') plt.legend() plt.xticks(index, labels) plt.ylim(0, 1) plt.show()
def main(): ap = argparse.ArgumentParser() ap.add_argument('--arch', required=True, nargs='+', type=int) ap.add_argument('--lr', required=False, type=int, default=.01) ap.add_argument('--epochs', required=False, type=int, default=100) ap.add_argument('--iters', required=True, type=int) ap.add_argument('--trials', required=True, type=int) ap.add_argument('--env', required=True) ap.add_argument('--t', required=True, type=int) ap.add_argument('--grads', required=True, type=int) opt = Options() opt.load_args(ap.parse_args()) opt.envname = opt.env opt.filename = '/Users/JonathanLee/experts/' + opt.envname + '.pkl' opt.env = gym.envs.make(opt.envname).env opt.sim = gym.envs.make(opt.envname).env opt.pi = net.Network([64, 64], opt.lr, opt.epochs) opt.pi.load_weights('meta/Pusher3-v0/weights_1.0-0.1-0.5_5.0.txt', 'meta/Pusher3-v0/stats_1.0-0.1-0.5_5.0.txt') opt.sess = None opt.sup = Supervisor(opt.pi) opt.misc = Options() opt.misc.num_evaluations = 1 plot_dir = utils.generate_plot_dir('initial', 'experts', vars(opt)) data_dir = utils.generate_data_dir('initial', 'experts', vars(opt)) if not os.path.exists(plot_dir): os.makedirs(plot_dir) if not os.path.exists(data_dir): os.makedirs(data_dir) if not os.path.exists(plot_dir + '/scores'): os.makedirs(plot_dir + '/scores') if not os.path.exists(plot_dir + '/mags'): os.makedirs(plot_dir + '/mags') opt.plot_dir = plot_dir opt.data_dir = data_dir sup_rewards = np.zeros((opt.trials, opt.misc.num_evaluations)) lnr_rewards = np.zeros((opt.trials, opt.misc.num_evaluations)) rob_rewards = np.zeros((opt.trials, opt.misc.num_evaluations)) train_err = np.zeros((opt.trials, opt.misc.num_evaluations, opt.t)) valid_err = np.zeros((opt.trials, opt.misc.num_evaluations, opt.t)) test_err = np.zeros((opt.trials, opt.misc.num_evaluations, opt.t)) robust_err = np.zeros((opt.trials, opt.misc.num_evaluations, opt.t)) freq = np.zeros((opt.trials, opt.misc.num_evaluations)) train_bar_errs = np.zeros((opt.trials, opt.t)) valid_bar_errs = np.zeros((opt.trials, opt.t)) test_bar_errs = np.zeros((opt.trials, opt.t)) print "Running Trials:\n\n" try: for t in range(opt.trials): start_time = timer.time() results = run_trial(opt) sup_rewards[t, :] = results['sup_reward'] lnr_rewards[t, :] = results['lnr_reward'] rob_rewards[t, :] = results['rob_reward'] train_err[t, :, :] = results['train_err'] valid_err[t, :, :] = results['valid_err'] test_err[t, :, :] = results['test_err'] robust_err[t, :, :] = results['robust_err'] freq[t, :] = results['correction_freq'] train_bar_errs[t], valid_bar_errs[t], test_bar_errs[t] = results[ 'bar_errs'] sup_rewards_save, lnr_rewards_save, rob_rewards_save = sup_rewards[: t + 1, :], lnr_rewards[: t + 1, :], rob_rewards[: t + 1, :] train_err_save, valid_err_save, test_err_save, robust_err_save = train_err[: t + 1, :, :], valid_err[: t + 1, :, :], test_err[: t + 1, :, :], robust_err[: t + 1, :, :] freq_save = freq[:t + 1, :] pd.DataFrame(sup_rewards_save).to_csv(opt.data_dir + 'sup_rewards.csv', index=False) pd.DataFrame(lnr_rewards_save).to_csv(opt.data_dir + 'lnr_rewards.csv', index=False) pd.DataFrame(rob_rewards_save).to_csv(opt.data_dir + 'rob_rewards.csv', index=False) for tau in range(opt.t): pd.DataFrame(train_err_save[:, :, tau]).to_csv( opt.data_dir + 'train_err_t' + str(tau) + '.csv', index=False) pd.DataFrame(valid_err_save[:, :, tau]).to_csv( opt.data_dir + 'valid_err_t' + str(tau) + '.csv', index=False) pd.DataFrame(test_err_save[:, :, tau]).to_csv( opt.data_dir + 'test_err_t' + str(tau) + '.csv', index=False) pd.DataFrame(robust_err_save[:, :, tau]).to_csv( opt.data_dir + 'robust_err_t' + str(tau) + '.csv', index=False) pd.DataFrame(freq_save).to_csv(opt.data_dir + 'freq.csv', index=False) train_err_avg = np.mean(train_err_save, axis=2) valid_err_avg = np.mean(valid_err_save, axis=2) test_err_avg = np.mean(test_err_save, axis=2) robust_err_avg = np.mean(robust_err_save, axis=2) utils.plot([sup_rewards_save, lnr_rewards_save, rob_rewards_save], ['Supervisor', 'Learner', 'Robust Learner'], opt, "Reward", colors=['red', 'blue', 'green']) utils.plot( [train_err_avg, valid_err_avg, test_err_avg, robust_err_avg], ['Training', 'Validation', 'Learner', 'Robust Learner'], opt, "Error", colors=['red', 'orange', 'blue', 'green']) utils.plot([freq_save], ['Frequency'], opt, 'Correction Frequency', colors=['green']) bar_errs = [ np.mean(train_bar_errs, axis=0), np.mean(valid_bar_errs, axis=0), np.mean(test_bar_errs, axis=0) ] labels = ['train', 'valid', 'test'] width = .2 index = np.arange(opt.t) for i, (err, label) in enumerate(zip(bar_errs, labels)): plt.bar(index + i * width, err, width, label=label) plt.legend() plt.ylim(0, .75) plt.savefig('/Users/JonathanLee/Desktop/bar_new_avg.png') utils.clear() end_time = timer.time() print "Trial time: " + str(end_time - start_time) except KeyboardInterrupt: pass
def main(): ap = argparse.ArgumentParser() ap.add_argument('--arch', required=True, nargs='+', type=int) ap.add_argument('--lr', required=False, type=float, default=.01) ap.add_argument('--epochs', required=False, type=int, default=100) ap.add_argument('--iters', required=True, type=int) ap.add_argument('--trials', required=True, type=int) ap.add_argument('--env', required=True) ap.add_argument('--t', required=True, type=int) ap.add_argument('--grads', required=True, type=int) ap.add_argument('--weights', required=True, nargs='+', type=float, default=[1.0, .1, .5]) ap.add_argument('--ufact', required=True, default=4.0, type=float) ap.add_argument('--id', required=True, default=4.0, type=int) ap.add_argument('--nu', required=True, type=float) ap.add_argument('--gamma', required=True, type=float) opt = Options() args = ap.parse_args() opt.load_args(args) args = vars(args) opt.envname = opt.env opt.env = gym.envs.make(opt.envname).env opt.sim = gym.envs.make(opt.envname).env exp_id = args['id'] opt.env.my_weights = args['weights'] opt.env.ufact = args['ufact'] opt.pi = net.Network([64, 64], .01, 300) suffix = '_' + utils.stringify(args['weights']) + '_' + str(args['ufact']) weights_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str( exp_id) + '_weights' + suffix + '.txt' stats_path = 'meta/' + 'test' + '/' + opt.envname + '_' + str( exp_id) + '_stats' + suffix + '.txt' opt.pi.load_weights(weights_path, stats_path) opt.sup = Supervisor(opt.pi) opt.misc = Options() opt.misc.num_evaluations = 1 opt.misc.samples = 200 rec_results = {} lnr_results = {} plot_dir = utils.generate_plot_dir('initial', 'experts', vars(opt)) data_dir = utils.generate_data_dir('initial', 'experts', vars(opt)) if not os.path.exists(plot_dir): os.makedirs(plot_dir) if not os.path.exists(data_dir): os.makedirs(data_dir) if not os.path.exists(plot_dir + '/scores'): os.makedirs(plot_dir + '/scores') if not os.path.exists(plot_dir + '/mags'): os.makedirs(plot_dir + '/mags') try: for t in range(opt.trials): start_time = timer.time() results = run_trial(opt) for key in results['rec'].keys(): if key in rec_results: rec_results[key].append(results['rec'][key]) else: rec_results[key] = [results['rec'][key]] for key in results['lnr'].keys(): if key in lnr_results: lnr_results[key].append(results['lnr'][key]) else: lnr_results[key] = [results['lnr'][key]] except KeyboardInterrupt: pass labels = sorted(list(rec_results.keys())) for key in rec_results.keys(): rec_results[key] = np.array(rec_results[key]) / float(opt.misc.samples) lnr_results[key] = np.array(lnr_results[key]) / float(opt.misc.samples) rec_means = [ np.mean(rec_results[key]) for key in sorted(rec_results.keys()) ] rec_sems = [ scipy.stats.sem(rec_results[key]) for key in sorted(rec_results.keys()) ] lnr_means = [ np.mean(lnr_results[key]) for key in sorted(lnr_results.keys()) ] lnr_sems = [ scipy.stats.sem(lnr_results[key]) for key in sorted(lnr_results.keys()) ]