mirror of
https://github.com/saymrwulf/stable-baselines3.git
synced 2026-11-17 08:56:08 +00:00
Use HardTanh to relax the constrain
This commit is contained in:
parent
c05c990285
commit
bff0ca0ea8
1 changed files with 3 additions and 1 deletions
|
|
@ -49,7 +49,9 @@ class Actor(BaseNetwork):
|
||||||
latent_sde_dim=latent_sde_dim,
|
latent_sde_dim=latent_sde_dim,
|
||||||
log_std_init=log_std_init)
|
log_std_init=log_std_init)
|
||||||
# Avoid saturation by limiting the mean of the gaussian to be in [-1, 1]
|
# Avoid saturation by limiting the mean of the gaussian to be in [-1, 1]
|
||||||
self.mu = nn.Sequential(self.mu, nn.Tanh())
|
# self.mu = nn.Sequential(self.mu, nn.Tanh())
|
||||||
|
# TODO: test with small positive slope to have non zero gradient
|
||||||
|
self.mu = nn.Sequential(self.mu, nn.Hardtanh(min_val=-2.0, max_val=2.0))
|
||||||
else:
|
else:
|
||||||
self.action_dist = SquashedDiagGaussianDistribution(action_dim)
|
self.action_dist = SquashedDiagGaussianDistribution(action_dim)
|
||||||
self.mu = nn.Linear(net_arch[-1], action_dim)
|
self.mu = nn.Linear(net_arch[-1], action_dim)
|
||||||
|
|
|
||||||
Loading…
Reference in a new issue