e07f9a24e60a8acc756efc0728adb257d0474c2f,catalyst/rl/agent/policy.py,RealNVPPolicy,forward,#RealNVPPolicy#Any#Any#Any#,76

Before Change

        )
        sigma = torch.ones_like(mu).to(mu.device)
        z = mu if deterministic else normal_sample(mu, sigma)
        log_pi = normal_logprob(mu, sigma, z)
        z, log_pi = self.coupling1.forward(z, state_embedding, log_pi)
        z, log_pi = self.coupling2.forward(z, state_embedding, log_pi)
        action, log_pi = self.squashing_layer.forward(z, log_pi)

After Change

        bool_logprob = isinstance(logprob, bool) and logprob
        value_logprob = isinstance(logprob, torch.Tensor)
        assert not value_logprob, "Not implemented behaviour"

        action_logprob = normal_logprob(loc, scale, action)
        action, action_logprob = \
            self.coupling1.forward(action, state_embedding, action_logprob)

Instances