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)
args = vars(ap.parse_args())
opt = Options()
opt.envname = args['env']
opt.trials = args['trials']
opt.iters = args['iters']
opt.epochs = args['epochs']
opt.lr = args['lr']
opt.arch = args['arch']
opt.t = args['t']
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 = load_policy.load_policy(opt.filename)
opt.sess = tf.Session()
opt.sup = NetSupervisor(opt.pi, opt.sess)
run_trial(opt)
def load_env(envname):
filename = '/Users/JonathanLee/experts/' + envname + '.pkl'
env = gym.envs.make(envname).env
pi = load_policy.load_policy(filename)
sess = tf.Session()
sup = NetSupervisor(pi, sess)
return env, sup
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)
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 = load_policy.load_policy(opt.filename)
opt.sess = tf.Session()
opt.sup = NetSupervisor(opt.pi, opt.sess)
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)
opt.plot_dir = plot_dir
opt.data_dir = data_dir
train_errs = np.zeros((opt.trials, opt.t))
valid_errs = np.zeros((opt.trials, opt.t))
test_errs = np.zeros((opt.trials, opt.t))
for t in range(opt.trials):
train_errs[t, :], valid_errs[t, :], test_errs[t, :] = run_trial(opt, t)
train_err = np.mean(train_errs, axis=0)
valid_err = np.mean(valid_errs, axis=0)
test_err = np.mean(test_errs, axis=0)
errs = [train_err, valid_err, test_err]
labels = ['train', 'valid', 'test']
width = .2
index = np.arange(opt.t)
for i, (err, label) in enumerate(zip(errs, labels)):
plt.bar(index + i * width, err, width, label=label)
plt.legend()
plt.ylim(0, .75)
plt.savefig('/Users/JonathanLee/Desktop/bar_original_avg.png')
utils.clear()
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import gym
import matplotlib.pyplot as plt
from icra_tools.supervisor import NetSupervisor
import tensorflow as tf
import numpy as np
from icra_tools.expert import load_policy
from icra_tools import statistics
import IPython
from sklearn import svm
from icra_tools import utils
envname = 'Walker2d-v1'
filename = '/Users/JonathanLee/experts/' + envname + '.pkl'
env = gym.envs.make(envname)
pi = load_policy.load_policy(filename)
sess = tf.Session()
sup = NetSupervisor(pi, sess)
T = 1000
ITERATIONS = 10
trajs_train = []
trajs_test = []
for i in range(7):
print "{ Iteration: " + str(i) + " }"
s = env.reset()
reward = 0.0
traj = []
for t in range(T):
# env.render()
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 = load_policy.load_policy(opt.filename)
opt.sess = tf.Session()
opt.sup = NetSupervisor(opt.pi, opt.sess)
opt.misc = Options()
opt.misc.num_evaluations = 10
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=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)
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 = load_policy.load_policy(opt.filename)
opt.sess = tf.Session()
opt.sup = NetSupervisor(opt.pi, opt.sess)
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)
opt.plot_dir = plot_dir
opt.data_dir = data_dir
sup_rewards = np.zeros((opt.trials, opt.iters))
lnr_rewards = np.zeros((opt.trials, opt.iters))
rob_rewards = np.zeros((opt.trials, opt.iters))
train_err = np.zeros((opt.trials, opt.iters))
valid_err = np.zeros((opt.trials, opt.iters))
test_err = np.zeros((opt.trials, opt.iters))
robust_err = np.zeros((opt.trials, opt.iters))
freq = np.zeros((opt.trials, opt.iters))
for t in range(opt.trials):
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']
pd.DataFrame(sup_rewards).to_csv(opt.data_dir + 'sup_rewards.csv',
index=False)
pd.DataFrame(lnr_rewards).to_csv(opt.data_dir + 'lnr_rewards.csv',
index=False)
pd.DataFrame(rob_rewards).to_csv(opt.data_dir + 'rob_rewards.csv',
index=False)
pd.DataFrame(train_err).to_csv(opt.data_dir + 'train_err.csv', index=False)
pd.DataFrame(valid_err).to_csv(opt.data_dir + 'valid_err.csv', index=False)
pd.DataFrame(test_err).to_csv(opt.data_dir + 'test_err.csv', index=False)
pd.DataFrame(robust_err).to_csv(opt.data_dir + 'robust_err.csv',
index=False)
pd.DataFrame(freq).to_csv(opt.data_dir + 'freq.csv', index=False)
utils.plot([sup_rewards, lnr_rewards, rob_rewards],
['Supervisor', 'Learner', 'Robust Learner'],
opt,
"Reward",
colors=['red', 'blue', 'green'])
utils.plot([train_err, valid_err, test_err, robust_err],
['Training', 'Validation', 'Learner', 'Robust Learner'],
opt,
"Error",
colors=['red', 'orange', 'blue', 'green'])
utils.plot([freq], ['Frequency'],
opt,
'Correction Frequency',
colors=['green'])