65713b1715b1628d99dac611a8853190b0c91f37,ml/rl/models/actor.py,GaussianFullyConnectedActor,forward,#GaussianFullyConnectedActor#Any#,143

Before Change

            // ONNX doesn"t like reshape either..
            return rlt.ActorOutput(action=action)
        // Since each dim are independent, log-prob is simply sum
        log_prob = torch.sum(
            self._log_prob(r, scale_log) - self._squash_correction(action), dim=1
        )
        return rlt.ActorOutput(action=action, log_prob=log_prob.reshape(-1, 1))

    def _atanh(self, x):
        

After Change

        if SummaryWriterContext._global_step % 1000 == 0:
            SummaryWriterContext.add_histogram("actor/forward/loc", loc.detach().cpu())
            SummaryWriterContext.add_histogram(
                "actor/forward/scale_log", scale_log.detach().cpu()
            )
            SummaryWriterContext.add_histogram(
                "actor/forward/log_prob", log_prob.detach().cpu()

Instances