cgca = data['trainer/cgca']
cactor_n = data['trainer/cactor_n']
from rlkit.torch.policies.make_deterministic import MakeDeterministic
cactor_n = [MakeDeterministic(cactor) for cactor in cactor_n]
for a2 in a2s:
o_n = env.reset()
cg_inputs = torch.cat([
torch.tensor(o_n),
torch.tensor([[np.random.rand() * 2. - 1.], [a2]])
],
dim=-1).float()[None, :]
# cg_inputs = torch.cat([torch.tensor(o_n), torch.tensor([[0.5],[a2]])],dim=-1).float()[None,:]
if context_graph:
contexts = context_graph(cg_inputs)
c1, _ = cactor.get_action(contexts[:, 0, :])
c1s.append(c1[0])
else:
if cactor:
cactions, _ = cactor.get_action(cg_inputs[0])
c1s.append(cactions[0, :])
else:
contexts = cgca(cg_inputs)
c1, _ = cactor_n[0].get_action(contexts[:, 0, :])
c1s.append(c1[0])
for a1 in a1s:
o_n = env.reset()
cg_inputs = torch.cat([
torch.tensor(o_n),
torch.tensor([[a1], [np.random.rand() * 2. - 1.]])
eval_policy = MakeDeterministic(policy)
returns = []
success_num = 0
collision_num = 0
inference_correct = 0
inference_total = 0
for _ in range(args.epoch):
o = env.reset()
policy.reset()
path_length = 0
done = False
c_r = 0.
while True:
path_length += 1
a, agent_info = eval_policy.get_action(o)
o, r, done, env_info = env.step(a)
if 'intentions' in agent_info.keys():
intention_probs = agent_info['intentions']
inffered_intentions = np.argmax(intention_probs,
axis=-1)
true_intentions = env.get_sup_labels()
valid_mask = ~np.isnan(true_intentions)
true_intentions = true_intentions[valid_mask]
inffered_intentions = inffered_intentions[valid_mask]
inference_correct += np.sum(
inffered_intentions == true_intentions)
inference_total += np.sum(valid_mask)
else:
inference_total += 1
if args.eval:
if not args.log_dir.startswith('MADDPG'):
from rlkit.torch.policies.make_deterministic import MakeDeterministic
p1 = MakeDeterministic(data['trainer/trained_policy_n'][0])
p2 = MakeDeterministic(data['trainer/trained_policy_n'][1])
plt.figure()
with torch.no_grad():
for _ in range(args.num):
o_n = env.reset()
xs = [o_n[0][0]]
ys = [o_n[0][1]]
obs1 = torch.tensor(o_n[0]).float()
obs2 = torch.tensor(o_n[1]).float()
for i in range(5):
a1, _ = p1.get_action(obs1)
a2, _ = p2.get_action(obs2)
o_n, r_n, done, info = env.step([a1, a2])
xs.append(o_n[0][0])
ys.append(o_n[0][1])
obs1 = torch.tensor(o_n[0]).float()
obs2 = torch.tensor(o_n[1]).float()
plt.plot(xs, ys, 'o-')
plt.plot(xs[0], ys[0], 'o', color='green')
plt.plot(xs[-1], ys[-1], 'o', color='black')
plt.gca().set_aspect('equal', 'box')
plt.xlim(-1, 1)
plt.ylim(-1, 1)
plt.xlabel('x')
plt.ylabel('y')
with torch.no_grad():
for agent in range(args.num_ag):
cs[agent] = []
if agent < int(args.num_ag/2):
for a2 in opponent_as:
o_n = env.reset()
actions = torch.zeros(args.num_ag)
actions[:int(args.num_ag/2)] = 0.
actions[int(args.num_ag/2):] = a2
actions = actions[:,None]
cg_inputs = torch.cat([torch.tensor(o_n), actions],dim=-1).float()[None,:]
if cactor:
if cgca:
contexts = cgca(cg_inputs)
cactions, _ = cactor.get_action(contexts[0])
else:
cactions, _ = cactor.get_action(cg_inputs[0])
cs[agent].append(cactions[agent,0])
else:
if cgca:
contexts = cgca(cg_inputs)
else:
contexts = cgca_n[agent](cg_inputs)
ca, _ = cactor_n[agent].get_action(contexts[:,agent,:][0])
cs[agent].append(ca[0])
else:
for a1 in opponent_as:
o_n = env.reset()
actions = torch.zeros(args.num_ag)
actions[:int(args.num_ag/2)] = a1
c2s = []
d_path = pre_path + '/' + 'seed' + str(args.seed) + '/params.pkl'
data = torch.load(d_path, map_location='cpu')
c1net = data['trainer/cactor_n'][0]
c2net = data['trainer/cactor_n'][1]
from rlkit.torch.policies.make_deterministic import MakeDeterministic
c1net = MakeDeterministic(c1net)
c2net = MakeDeterministic(c2net)
for a2 in a2s:
o_n = env.reset()
c1_input = torch.tensor([o_n[0][0], a2]).float()
c1, _ = c1net.get_action(c1_input)
c1s.append(c1[0])
for a1 in a1s:
o_n = env.reset()
c2_input = torch.tensor([o_n[1][0], a1]).float()
c2, _ = c2net.get_action(c2_input)
c2s.append(c2[0])
# plt.figure()
# plt.subplot(1,2,1)
# plt.plot(a2s,c1s)
# plt.gca().set_aspect('equal', 'box')
# plt.gca().axes.get_xaxis().set_ticks([-1.,0.,1.])
# plt.gca().axes.get_yaxis().set_ticks([-1.,0.,1.])
# plt.xlim(-1,1)