Why tf.multinomial works with negative inputs like [[-1,1]]?

Question

I guess tf.multinomial does sampling based on softmax([[-1,1]]). If you have some ideas, please, tell me.

Answer 1

That's because tf.multinomial takes log-probabilities as its first argument.

From the docs:

logits: 2-D Tensor with shape [batch_size, num_classes]. Each slice [i, :] represents the unnormalized log-probabilities for all classes.