# Recurrence
if args.commnet and (args.recurrent or args.rnn_type == 'LSTM'):
args.recurrent = True
args.rnn_type = 'LSTM'
parse_action_args(args)
if args.seed == -1:
args.seed = np.random.randint(0, 10000)
torch.manual_seed(args.seed)
#print(args)
if args.commnet:
policy_net = CommNetMLP(args, num_inputs)
elif args.random:
policy_net = Random(args, num_inputs)
elif args.recurrent:
policy_net = RNN(args, num_inputs)
else:
policy_net = MLP(args, num_inputs)
if not args.display:
display_models([policy_net])
# share parameters among threads, but not gradients
for p in policy_net.parameters():
p.data.share_memory_()
if args.nprocesses > 1:
parse_action_args(args)
if args.seed == -1:
args.seed = np.random.randint(0, 10000)
torch.manual_seed(args.seed)
print(args)
if args.gacomm:
policy_net = GACommNetMLP(args, num_inputs)
elif args.commnet:
if args.tarcomm:
policy_net = TarCommNetMLP(args, num_inputs)
else:
policy_net = CommNetMLP(args, num_inputs)
elif args.random:
policy_net = Random(args, num_inputs)
elif args.recurrent:
policy_net = RNN(args, num_inputs)
else:
policy_net = MLP(args, num_inputs)
if not args.display:
display_models([policy_net])
# share parameters among threads, but not gradients
for p in policy_net.parameters():
p.data.share_memory_()
if args.env_name == 'grf':