Reversible Graph to Seqeuence model

refrence model

• Encoder: Graph2Seq

• Decoder: Semi-Supervised Neural Architecture Search

• Dataset :

  • Nasbench101 : paper
  • Nasbench101 : code

Encoder

• Input : Graph (Adjacency matrix, list of operation)
• output : sequence

For graph G,

1. Create one virtual node called supernode S.
   We define embedding(encoded sequence) of G as the node embedding of supernode S
   Add edges from every node in G to Supernode S.
2. Expand G into sequence.(Follow semiNAS paper method)
3. Loop node embedding : aggregate neighbor nodes.(Follow Graph2Seq paper method)
4. Node embedding of supernode S at the last step is an embedding(encoded sequence) of G

Decoder

• input: sequence
• ouput : Graph (Adjacency matrix, list of operation)