I have read the paper "Learning to Read by Spelling" by Gutpa et al.
They present a method for visual text recognition without using any paired supervisory data. In chapter 4 they describe how to implement their method.
Now comes the part I didn't quite understand for the (Text-)Recogniser:
After an image has passed through 4 conv blocks, you get an output of (2, n, D) (n = number of maximum character length, D = 32 = number of features). Then average pooling is applied and you get an output of (n, D). Each of the n D-dimensional feature vectors are mapped to |A| = K dimensional logits through linear projection, yielding a K × n dimensional tensor. |A| = K is the alphabet A containing K symbols.
From my understanding, the K × n dimensional tensor is one-hot-encoded for the symbols of the alphabet A. How can you map such a tensor to an alphabet?
Thanks for the help.