def test_policy_opt_live():
test_dir = gps_path + '/tests/tests_tensorflow' + '/test_data/'
obs = np.load(test_dir + 'obs.npy')
tgt_mu = np.load(test_dir + 'tgt_mu.npy')
tgt_prc = np.load(test_dir + 'tgt_prc.npy')
scale = np.load(test_dir + 'scale_npy.npy')
bias = np.load(test_dir + 'bias_npy.npy')
hyper_params = POLICY_OPT_TF
deg_obs = 4
deg_action = 2
policy = PolicyOptTf(hyper_params, deg_obs, deg_action)
policy.policy.scale = scale
policy.policy.bias = bias
iterations = 200
batch_size = 32
batches_per_epoch = np.floor(800 / batch_size)
idx = range(800)
np.random.shuffle(idx)
for i in range(iterations):
# Load in data for this batch.
start_idx = int(i * batch_size %
(batches_per_epoch * batch_size))
idx_i = idx[start_idx:start_idx+batch_size]
feed_dict = {policy.obs_tensor: obs[idx_i],
policy.action_tensor: tgt_mu[idx_i],
policy.precision_tensor: tgt_prc[idx_i]}
t = policy.sess.run(policy.act_op, feed_dict={policy.obs_tensor: np.expand_dims(obs[idx_i][0], 0)})
policy.solver(feed_dict, policy.sess)
def test_load_from_auto_save():
import pickle
path_to_dict = gps_path + '/gps/algorithm/policy_opt/tf_checkpoint/policy_checkpoint.ckpt_hyperparams'
state = pickle.load(open(path_to_dict, "rb"))
hyper_params = state['hyperparams']
deg_obs = state['dO']
deg_action = state['dU']
policy_opt = PolicyOptTf(hyper_params, deg_obs, deg_action)
policy_opt.__setstate__(state)
def test_policy_opt_tf_backwards():
hyper_params = POLICY_OPT_TF
deg_obs = 14
deg_action = 7
policy_opt = PolicyOptTf(hyper_params, deg_obs, deg_action)
N = 20
T = 30
obs = np.random.randn(N, T, deg_obs)
obs_reshaped = np.reshape(obs, (N*T, deg_obs))
policy_opt.policy.scale = np.diag(1.0 / np.std(obs_reshaped, axis=0))
policy_opt.policy.bias = -np.mean(obs_reshaped.dot(policy_opt.policy.scale), axis=0)
policy_opt.prob(obs=obs)
def test_policy_opt_backwards():
hyper_params = POLICY_OPT_TF
deg_obs = 14
deg_action = 7
policy_opt = PolicyOptTf(hyper_params, deg_obs, deg_action)
# pylint: disable=W0212
policy_opt._hyperparams['iterations'] = 100 # 100 for testing.
N = 10
T = 10
obs = np.random.randn(N, T, deg_obs)
tgt_mu = np.random.randn(N, T, deg_action)
tgt_prc = np.random.randn(N, T, deg_action, deg_action)
tgt_wt = np.random.randn(N, T)
new_policy = policy_opt.update(obs, tgt_mu, tgt_prc, tgt_wt, itr=0, inner_itr=1)
def test_unpickle():
hyper_params = POLICY_OPT_TF
deg_obs = 14
deg_action = 7
policy_opt = PolicyOptTf(hyper_params, deg_obs, deg_action)
N = 20
T = 30
obs = np.random.randn(N, T, deg_obs)
obs_reshaped = np.reshape(obs, (N*T, deg_obs))
scale = np.diag(1.0 / np.std(obs_reshaped, axis=0))
bias = -np.mean(obs_reshaped.dot(scale), axis=0)
hyper_params['scale'] = scale
hyper_params['bias'] = bias
hyper_params['tf_iter'] = 100
policy_opt.__setstate__({'hyperparams': hyper_params, 'dO': deg_obs, 'dU': deg_action,
'scale': policy_opt.policy.scale, 'bias': policy_opt.policy.bias, 'tf_iter': 100})
def test_pickle():
hyper_params = POLICY_OPT_TF
deg_obs = 100
deg_action = 7
policy_opt = PolicyOptTf(hyper_params, deg_obs, deg_action)
state = policy_opt.__getstate__()
def test_auto_save_state():
hyper_params = POLICY_OPT_TF
deg_obs = 100
deg_action = 7
policy_opt = PolicyOptTf(hyper_params, deg_obs, deg_action)
policy_opt.auto_save_state()
def test_auto_save_state():
hyper_params = POLICY_OPT_TF
deg_obs = 100
deg_action = 7
policy_opt = PolicyOptTf(hyper_params, deg_obs, deg_action)
policy_opt.auto_save_state()
def test_pickle():
hyper_params = POLICY_OPT_TF
deg_obs = 100
deg_action = 7
policy_opt = PolicyOptTf(hyper_params, deg_obs, deg_action)
state = policy_opt.__getstate__()
def test_policy_opt_tf_init():
hyper_params = POLICY_OPT_TF
deg_obs = 100
deg_action = 7
PolicyOptTf(hyper_params, deg_obs, deg_action)
