diff --git a/torchy_baselines/sac/policies.py b/torchy_baselines/sac/policies.py index 58d5a99..9de90e6 100644 --- a/torchy_baselines/sac/policies.py +++ b/torchy_baselines/sac/policies.py @@ -76,7 +76,7 @@ class Actor(BaseNetwork): 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.Hardtanh(min_val=-2.0, max_val=2.0)) + # self.mu = nn.Sequential(self.mu, nn.Hardtanh(min_val=-2.0, max_val=2.0)) # Small positive slope to have non-zero gradient self.mu = nn.Sequential(self.mu, LeakyClip()) else: