why do we need replace_batchnorm ?

[TranThanh96]

why do you need replace_batchnorm when eval?

[Isalia20]

Not the author of the repo but I assume it's done to achieve maximum performance on IoT devices. Batch norm is fused with convolution operation and is done with just convolution(by fusing the conv weights with batchnorm stats+affine transforms). It first fuses batch norm with convolution and then replaces batch norm layer with identity layer(multiply by 1 basically) to speed up the computation

[jameslahm]

As @Isalia20 explains. Thanks!