def chunked_forwarding()

in evaluate.py [0:0]

• 31 lines of code
• 6 McCabe index (conditional complexity)

def chunked_forwarding(net, mono, view):
    '''
    binauralized the mono input given the view
    :param net: binauralization network
    :param mono: 1 x T tensor containing the mono audio signal
    :param view: 7 x K tensor containing the view as 3D positions and quaternions for orientation (K = T / 400)
    :return: 2 x T tensor containing binauralized audio signal
    '''
    net.eval().cuda()
    mono, view = mono.cuda(), view.cuda()

    chunk_size = 480000  # forward in chunks of 10s
    rec_field = net.receptive_field() + 1000  # add 1000 samples as "safe bet" since warping has undefined rec. field
    rec_field -= rec_field % 400  # make sure rec_field is a multiple of 400 to match audio and view frequencies
    chunks = [
        {
            "mono": mono[:, max(0, i-rec_field):i+chunk_size],
            "view": view[:, max(0, i-rec_field)//400:(i+chunk_size)//400]
        }
        for i in range(0, mono.shape[-1], chunk_size)
    ]

    for i, chunk in enumerate(chunks):
        with th.no_grad():
            mono = chunk["mono"].unsqueeze(0)
            view = chunk["view"].unsqueeze(0)
            binaural = net(mono, view)["output"].squeeze(0)
            if i > 0:
                binaural = binaural[:, -(mono.shape[-1]-rec_field):]
            chunk["binaural"] = binaural

    binaural = th.cat([chunk["binaural"] for chunk in chunks], dim=-1)
    binaural = th.clamp(binaural, min=-1, max=1).cpu()
    return binaural