Skip to contents

Tracing refuses a captured tensor that carries a gradient, and the fitted weights all do, so they are switched off around the recording and switched back afterwards. Nothing else about the network changes.

Usage

mdn_trace(de, xt, tt)