saymrwulf/stable-baselines3
1
0
Fork 0
mirror of https://github.com/saymrwulf/stable-baselines3.git synced 2026-05-31 23:28:05 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Use HardTanh to relax the constrain

Browse source
This commit is contained in:
Antonin Raffin 2019-12-19 11:59:00 +01:00
parent c05c990285
commit bff0ca0ea8
1 changed files with 3 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
torchy_baselines/sac/policies.py
View file

@ -49,7 +49,9 @@ class Actor(BaseNetwork):
latent_sde_dim=latent_sde_dim,
log_std_init=log_std_init)
# 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:
self.action_dist = SquashedDiagGaussianDistribution(action_dim)
self.mu = nn.Linear(net_arch[-1], action_dim)