def build_a2c_net_from_file(cfg, use_grad_cam=True):
net_file = cfg.A2C_FN
net = common.getNet(cfg)
net.load_state_dict(
torch.load(net_file, map_location=lambda storage, loc: storage))
net.eval()
net.grad_cam = use_grad_cam
return net
def build_a2c_nets_from_files(cfg, use_grad_cam=True):
net_files = [
cfg.I2A_RP_FN1, cfg.I2A_RP_FN2, cfg.I2A_RP_FN3, cfg.I2A_RP_FN4
]
nets = []
for file in range(4):
net = common.getNet(cfg)
net.load_state_dict(
torch.load(net_files[file],
map_location=lambda storage, loc: storage))
net.eval()
net.grad_cam = use_grad_cam
nets.append(net)
return nets
saves_path = writer.logdir
envs = [
common.makeCustomizedGridEnv(config) for _ in range(config.NUM_ENVS)
]
test_env = common.makeCustomizedGridEnv(config)
#sets seed on torch operations and on all environments
common.set_seed(config.SEED, envs=envs)
common.set_seed(config.SEED, envs=[test_env])
obs_shape = envs[0].observation_space.shape
act_n = envs[0].action_space.n
# net_policy = common.AtariA2C(obs_shape, act_n).to(device)
net_policy = common.getNet(config)
config.A2CNET = str(net_policy)
net_em = models.environment_model.EnvironmentModel(obs_shape, act_n,
config)
# net_em.load_state_dict(torch.load(config.EM_FILE_NAME, map_location=lambda storage, loc: storage))
net_em = net_em.to(device)
config.EM_NET = str(net_em)
net_i2a = i2a_model_no_LSTM.I2A_FC(obs_shape, act_n, net_em, net_policy,
config).to(device)
config.I2A_NET = str(net_i2a)
config.ROLLOUT_ENCODER = str(net_i2a.encoder)
# net_i2a.load_state_dict(torch.load("saves/03_i2a_test/best_pong_-018.667_1300.dat", map_location=lambda storage, loc: storage))
# print(net_policy)
# print(net_em)
help="learning rate")
fig, _ = plt.subplots()
config = ExperimentCfg()
config.make_test_env_config(parser)
device = torch.device(config.DEVICE)
env = common.makeCustomizedGridEnv(config)
device = torch.device("cuda")
config.DEVICE = 'cuda'
obs_shape = env.observation_space.shape
act_n = env.action_space.n
net = common.getNet(config)
net.load_state_dict(
torch.load(config.A2C_FN, map_location=lambda storage, loc: storage))
agent = ptan.agent.PolicyAgent(
lambda x: net(x)[0],
action_selector=ptan.actions.ProbabilityActionSelector(),
apply_softmax=True,
device=device)
state = env.reset()
total_rw = 0
total_steps = 0
episodes = config.EPISODES
for i in tqdm(range(episodes)):
"--INPUT",
default=False,
required=False,
help="")
config = ExperimentCfg()
config.make_replay_config(parser)
device = torch.device(config.DEVICE)
env = common.makeCustomizedGridEnv(config)
device = torch.device("cuda")
print(config.REPLACEMENT)
obs_shape = env.observation_space.shape
act_n = env.action_space.n
#load the a2c policy used for the EM training
net = common.getNet(device, config)
if (config.A2C_FN != ""):
net.load_state_dict(
torch.load(config.A2C_FN,
map_location=lambda storage, loc: storage))
agent = ptan.agent.PolicyAgent(
lambda x: net(x)[0],
action_selector=ptan.actions.ProbabilityActionSelector(),
apply_softmax=True,
device=device)
net_em = environment_model.EnvironmentModel(obs_shape, act_n, config)
net_em.load_state_dict(
torch.load(config.EM_FN, map_location=lambda storage, loc: storage))
net_em = net_em.to(device)
negativeReward=-5,
positiveReward=1)
return env
config = ExperimentCfg()
config.FRAME_SIZE = 14
device = torch.device(config.DEVICE)
env = makeCustomizedGridEnv()
device = torch.device('cpu')
obs_shape = env.observation_space.shape
act_n = env.action_space.n
net_a2c = common.getNet(config)
net_a2c.load_state_dict(
torch.load(
"//home/valy/OneDrive/repos/I2A-all/master/no-repl/5_14/Jan31_15-57-28_valy_a2c_14_5_0.0008_0.0001_False/best_0004.000_6000.dat",
map_location=lambda storage, loc: storage))
net_a2c.eval()
net_em = environment_model.EnvironmentModel(obs_shape, act_n, config)
net_em.load_state_dict(
torch.load(
"/home/valy/OneDrive/repos/I2A-all/master/no-repl/5_14/Jan31_16-40-11_valy_em_14_5_0.0008_0.0001_False/best_9.1968e-08_223449.dat",
map_location=lambda storage, loc: storage))
net_em = net_em.to(device)
net_em.eval()
net_distilled_policy = common.getNet(config)
parser.add_argument("-e", "--EPISODES", default=5000, type=int,required=False, help="")
parser.add_argument("-p", "--PLOT", default=False, required=False, help="")
parser.add_argument("-in", "--INPUT", default=False, required=False, help="")
parser.add_argument("-lr", required=True, type=float, help="learning rate")
config = ExperimentCfg()
config.make_i2a_replay_config(parser)
device = torch.device(config.DEVICE)
env = common.makeCustomizedGridEnv(config)
device = torch.device("cuda")
obs_shape = env.observation_space.shape
act_n = env.action_space.n
net_a2c = common.getNet(config)
net_a2c.load_state_dict(torch.load(config.A2C_FN, map_location=lambda storage, loc: storage))
net_a2c.eval()
net_a2c.to(device)
# if(config.IS_I2A):
net_em = environment_model.EnvironmentModel(obs_shape, act_n, config)
net_em.load_state_dict(torch.load(config.EM_FN, map_location=lambda storage, loc: storage))
net_em = net_em.to(device)
net_em.eval()
net_i2a = i2a_model.I2A(obs_shape, act_n, net_em, net_a2c, config).to(device)
net_i2a.load_state_dict(torch.load(config.I2A_FN, map_location=lambda storage, loc: storage))
net_i2a.eval()
# net = net_i2a