0ac2b33e8c63304a50db7d2b484368299706b58b,slm_lab/agent/net/mlp.py,MLPNet,training_step,#MLPNet#Any#Any#Any#Any#Any#,130

Before Change

        More most RL usage, we have custom, often complicated, loss functions. Compute its value and put it in a pytorch tensor then pass it in as loss
        """
        self.train()
        self.zero_grad()
        self.optim.zero_grad()
        if loss is None:
            out = self(x)
            loss = self.loss_fn(out, y)

After Change

        Takes a single training step: one forward and one backwards pass
        More most RL usage, we have custom, often complicated, loss functions. Compute its value and put it in a pytorch tensor then pass it in as loss
        """
        if hasattr(self, "model_tails") and x is not None:
            raise ValueError("Loss computation from x,y not supported for multitails")
        self.lr_scheduler.step(epoch=lr_t)
        self.train()
        self.optim.zero_grad()
        if loss is None:

Instances