What exactly Flux.destructure does?

Just to close this up, the problem was caused by nontrainable parameters in BatchNorm. Flux.params contains only trainable parameters, so the nontrainbale ones were not copied.