Closed nowittynamesleft closed 2 months ago
Added python notebook for acc. proving but proof step hangs
separated blocks of steps for proof and timed each step in c0b2955 . Step still hangs, but no real CPU usage during the proof step.
Fixed prove step hanging by calibrating with real data (binary tensors instead of real-valued random tensors). Input tensor shape is important. Split data into chunks with fixed batch size before proving & verified each proof
Need to make a proof of inference on test set, then prove the accuracy on the test set as well, and verify these for each merged model