gym.make(run.get("defaults", "env")))),
width=im_width,
height=im_height,
grayscale=grayscale)
if args.monitor:
e = gym.wrappers.Monitor(e, args.monitor)
return e
env = make_env()
env_pool = [env]
for i in range(run.getint("defaults", "env_pool_size", fallback=1) - 1):
env_pool.append(make_env())
params = env_params.EnvParams.from_env(env)
params.load_runfile(run)
env_params.register(params)
model = Net(params.n_actions,
input_shape=(1 if grayscale else 3, im_height, im_width))
if params.cuda_enabled:
model.cuda()
loss_fn = nn.MSELoss(size_average=False)
optimizer = optim.Adam(model.parameters(),
lr=run.getfloat("learning", "lr"))
action_selector = ActionSelectorEpsilonGreedy(epsilon=run.getfloat(
"defaults", "epsilon"),
params=params)
target_net = agent.TargetNet(model)
dqn_agent = agent.DQNAgent(dqn_model=model,
def test_register(self):
self.assertIsNone(env_params.get())
params = env_params.EnvParams.from_env(self.env)
env_params.register(params)
self.assertEqual(params, env_params.get())