Version 1
y = episode_a.argmax(-1) # episode_a is in shape [T, n_actions]
action_preds = self.net(ep_s) # action_preds is logits before softmax
neg_log_like = self.loss_fn(action_preds, y)
loss = torch.mean(r * neg_log_like) # r is reward
Version 2
y = torch.tensor(episode_a, requires_grad=True)
action_preds = model(ep_s)
neg_log_like = -y * torch.log(action_preds)
loss = torch.sum(neg_log_like, 1).mean()
Versions 1 and 2 seem to have same loss value. The difference is, y does not require grad in version while it does in version 2. But it is like a supervised learning backprop operation, y should be not needing require_grad. I do not understand why version 1 cannot train up the policy but version 2 can?