def get_stochastic_action(self, x):
with torch.no_grad():
assert len(x.shape) == 2
q1, q2 = self.forward(x)
probs = self.get_mean_distribution_from_Qs(q1, q2)
action = Categorical(probs).sample()
assert action.shape == (1, )
return action.detach().cpu().numpy()[0]
def select_actions(pi, dist_type, env_type):
if env_type == 'atari':
actions = Categorical(pi).sample()
else:
if dist_type == 'gauss':
mean, std = pi
actions = Normal(mean, std).sample()
elif dist_type == 'beta':
alpha, beta = pi
actions = Beta(alpha.detach().cpu(), beta.detach().cpu()).sample()
# return actions
return actions.detach().cpu().numpy().squeeze()
def select_actions(pi):
actions = Categorical(pi).sample()
# return actions
return actions.detach().cpu().numpy().squeeze()