def test(args=get_args()):
torch.set_num_threads(1) # we just need only one thread for NN
env = gym_make()
model_path = os.path.join(args.logdir, args.task, 'ddpg/policy.pth')
layer = [1024, 512, 512, 512]
device = 'cuda'
state_shape = env.observation_space.shape or env.observation_space.n
action_shape = env.action_space.shape or env.action_space.n
action_range = [env.action_space.low, env.action_space.high]
actor = Actor(
layer, state_shape, action_shape,
action_range, device
).to(device)
critic = Critic(
layer, state_shape, action_shape, device
).to(device)
actor = actor.to(device)
actor_optim = torch.optim.Adam(actor.parameters())
critic = critic.to(device)
critic_optim = torch.optim.Adam(critic.parameters())
policy = DDPGPolicy(
actor, actor_optim, critic, critic_optim,
action_range=action_range)
policy.load_state_dict(torch.load(model_path, map_location=device))
obs = env.reset()
# env.state[0] = -30.0
# env.goal[0] = 30.0
env.render()
print(env.goal)
while True:
action, _ = policy.actor(obs.reshape(1,-1), eps=0.01)
action = action.detach().cpu().numpy()[0]
obs, reward, done, info = env.step(action)
# print(env.state)
# print(reward)
# print(info)
env.render()
if done:
break
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
# create the objects for two networks and for the two optimizers
generator = Generator(latent=opt.latent, channels=opt.channels, num_filters=opt.num_filters)
critic = Critic(channels=opt.channels, num_filters=opt.num_filters)
optimizer_G = torch.optim.RMSprop(generator.parameters(), lr=opt.learning_rate)
optimizer_C = torch.optim.RMSprop(critic.parameters(), lr=opt.learning_rate)
# put the nets on gpu
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
generator, critic = generator.to(device), critic.to(device)
generator.apply(weights_init)
critic.apply(weights_init)
if opt.dataset == 'cifar10':
print(ROOT_DIR + "/cifar")
if not os.path.isdir(ROOT_DIR + "/cifar"):
os.mkdir(ROOT_DIR + "/cifar")
elif opt.dataset == 'LSUN':
if not os.path.isdir(ROOT_DIR + "/bedrooms"):
os.mkdir(ROOT_DIR + "/bedrooms")
# start training
current_epoch = 0
gen_iterations = 0
for epoch in range(opt.n_epochs):