Why does the funcFlopCompute run twice for each layer ?

[jiajunsong811]

def _wrapFunc(func, funcFlopCompute):
    global cuda_sync
    oldFunc = func
    name = func.__str__
    old_functions[name] = oldFunc
    func_name = func.__name__
    def newFunc(*args, **kwds):
        flops, macs = funcFlopCompute(*args, **kwds)
        if module_flop_count:
            module_flop_count[-1].append((name, flops))
        if module_mac_count and macs:
            module_mac_count[-1].append((name, macs))
        if cuda_sync:
            cuda_start = torch.cuda.Event(enable_timing=True)
            cuda_end = torch.cuda.Event(enable_timing=True)
            cuda_start.record()
        else:
            start = time.time()
        ret = oldFunc(*args, **kwds)
        if cuda_sync:
            cuda_end.record()
            torch.cuda.synchronize()
            duration = cuda_start.elapsed_time(cuda_end)/1000
        else:
            duration = time.time() - start
        if module_profile_lists:
            module_profile_lists[-1].append((func_name, profileEntry(flops, macs, duration)))
        return ret

    newFunc.__str__ = func.__str__

    return newFunc

For each layer, the funcFlopCompute function will run twice. This first run is the F.linear (or other F function). The second is "ret = oldFunc(*args, **kwds)". That make the flops is twice than the true flops.

Why does the funcFlopCompute run twice for each layer ?