@kinianlo Hi and thanks for the PR. In general, I think it's better to avoid any hidden manipulation at the tensor network level, since physical changes of this form can alter dramatically training. Perhaps a better solution here would be a spider-specific tensor ansatz which arranges the spiders reader output e.g. as follows, producing a chain of 3-legged spiders.

w1    w2
  \  /
   o      w3 
     \   /
       o    ....
        \   /
          o
          |

• Have a look at MPSAnsatz, SpiderAnsatz, which break large tensors similarly in generic string diagrams.
• Note this can be also implemented as a separate reader instead of a new ansatz, e.g. SpiderChainReader (perhaps this is even better).