def bellman_test_features(args):
if args.model != 'bellman':
raise Exception, 'wrong model for this'
(feature_fn, margin_fn, num_features, actions) = select_feature_fn(args)
demofile = h5py.File(args.demofile, 'r')
# get a random set of trajectories
trajectories = []
traj = []
for uid in range(len(demofile)):
key = str(uid)
group = demofile[key]
state = [key,group['cloud_xyz'][:]] # these are already downsampled
action = group['action'][()] if not group['knot'][()] else 'done'
if group['pred'][()] == key:
if traj:
trajectories.append(traj)
traj = []
traj.append([state, action])
if traj:
trajectories.append(traj)
random.shuffle(trajectories)
constraint_trajs = trajectories[:args.num_constraints]
# put them into a Bellman model
mm_model = BellmanMaxMarginModel(action, 500, 1000, 10, .9, num_features, feature_fn, margin_fn)
for i, t in enumerate(constraint_trajs):
mm_model.add_trajectory(t, str(i), True)
weights, w0 = mm_model.optimize_model()
# evaluate value fn performance
num_evals = len(trajectories) if args.num_evals < 0 else args.num_constraints + args.num_evals
trajectories = trajectories[args.num_constraints:num_evals]
values = np.zeros((len(trajectories), 6))
num_decreases = 0
for (i, t) in enumerate(trajectories):
for j, (s, a) in enumerate(t):
values[i, j] = np.dot(mm_model.weights, mm_model.feature_fn(s, a))
for k in range(j):
if values[i, k] > values[i, k+1]:
num_decreases += 1
print "DECREASE", values[i, k], values[i, k+1]
sys.stdout.write('num decreases:\t{} computed values for trajectory {}\r'.format(num_decreases, i))
sys.stdout.flush()
sys.stdout.write('\n')
print "num decreases:\t", num_decreases
ipy.embed()
