PyTorch currently has torch.nn.utils.fuse_conv_bn_eval (and I think some ConvBnRelu fusion with kernels only in quantized setting) which you must call manually (for details look at the source code, it’s really simple code), but it’s not documented yet
2 Likes