parser = argparse.ArgumentParser()
parser.add_argument("-i",
"--ini",
required=True,
help="Ini file to use for this run")
parser.add_argument("-n", "--name", required=True, help="Name of the run")
args = parser.parse_args()
config = conf.Config(args.ini)
device = torch.device("cuda" if config.train_cuda else "cpu")
name = config.train_name(suffix=args.name)
writer = SummaryWriter(comment="-" + name)
save_path = os.path.join("saves", name)
os.makedirs(save_path)
cube_env = cubes.get(config.cube_type)
assert isinstance(cube_env, cubes.CubeEnv)
log.info("Selected cube: %s", cube_env)
value_targets_method = model.ValueTargetsMethod(
config.train_value_targets_method)
net = model.Net(cube_env.encoded_shape,
len(cube_env.action_enum)).to(device)
print(net)
opt = optim.Adam(net.parameters(), lr=config.train_learning_rate)
sched = scheduler.StepLR(opt, 1, gamma=config.train_lr_decay_gamma
) if config.train_lr_decay_enabled else None
step_idx = 0
buf_policy_loss, buf_value_loss, buf_loss = [], [], []
buf_policy_loss_raw, buf_value_loss_raw, buf_loss_raw = [], [], []
parser.add_argument("-e",
"--env",
required=True,
help="Type of env to train, supported types=%s" %
cubes.names())
parser.add_argument("-m",
"--model",
required=True,
help="Model file to load")
parser.add_argument("-o",
"--output",
required=True,
help="Output prefix for plots")
args = parser.parse_args()
cube_env = cubes.get(args.env)
log.info("Selected cube: %s", cube_env)
net = model.Net(cube_env.encoded_shape, len(cube_env.action_enum))
net.load_state_dict(
torch.load(args.model, map_location=lambda storage, loc: storage))
net.eval()
log.info("Network loaded from %s", args.model)
# model.make_train_data(cube_env, net, device='cpu', batch_size=10, scramble_depth=2, shuffle=False)
states_by_depth = gen_states(cube_env,
max_depth=MAX_DEPTH,
round_counts=ROUND_COUNTS)
# for idx, states in enumerate(states_by_depth):
# log.info("%d: %s", idx, states)