embedding_layer: add more comments

Author: OneAdder

src/nf/nf_embedding_layer.f90

@@ -58,6 +58,7 @@ pure module subroutine backward(self, input, gradient)
     end subroutine backward
 
     pure module subroutine positional_encoding(self, pos)
+      !! Sum embedding with positional info (trigonometric, not trianable)
       class(embedding_layer), intent(in out) :: self
       integer, intent(in) :: pos
     end subroutine positional_encoding

src/nf/nf_layer_constructors.f90

@@ -234,6 +234,14 @@ module function self_attention(num_heads) result(res)
     end function self_attention
 
     module function embedding(sequence_length, vocab_size, model_dimension) result(res)
+      !! Embedding layer constructor.
+      !!
+      !! This layer is for inputting token indices from the dictionary to the network.
+      !! Works as a trainable lookup table that converts each index into a vector.
+      !! Embedding layer must be the first layer in a network.
+      !! `sequence_length`: max len of input sequence
+      !! `vocab_size`: length of token vocabulary
+      !! `model_dimension`: size of target embeddings
       integer, intent(in) :: sequence_length, vocab_size, model_dimension
       type(layer) :: res
     end function embedding