I really don't quite understand what the logits means in a2c.py

[Huixxi]

Thanks your great work, I was reading your amazing blog recently. Maybe a stupid issue, I don't really understand the logits means in your code. I only know that it is the raw output of the last Dense layer. But how can it be put into the tf.random.categorical(logits, 1) directly without any preprocessing? I mean it at least should be pass into a softmax layer to convert it to a probability distribution right? Or the softmax is an inner operation of tf.random.categorical(logits, 1) and so that we can pass the logits directly into that function to pick an action based on its probability? I tried to track its source code but failed.
Another question, what the logits means, q-values of each action or just their probability? As far as I know that layer should be a policy-based operation, so ... I think there is nothing to do with q-value or value-function.

[inoryy]

Generally, the logits are simply unnormalized log probabilities.

Mathematically, you would want to work with probability distributions in most cases, but numerically it often makes more sense to work directly with the logits - not only because you do fewer ops, but also because it can be more computationally efficient under the hood (e.g. random sampling from generic categorical distribution is usually implemented with log of probabilities anyway).

Specifically, indeed tf.random.categorical works directly with logits. You can read the python source code here, though it eventually traces back to internal bindings.

[Huixxi]

Yes, thanks for your reply, I tried to track their source code from the same link you provided. But I failed at line 392. So I still can't get how they deal with the raw logits. Now I just assume that tf.random.categorical has the convert logits to probability inner operation.

[Huixxi]

Another question, I think there is no need to x = tf.convert_to_tensor(inputs, dtype=tf.float32), I've checked that the input is already a tensor(but I don't know how it does), so I comments it and the code works well.