eval_rank.py 文件源码

def build_model(tparams, options):
    """
    Construct computation graph for the whole model
    """
    # inputs (image, sentence, contrast images, constrast sentences)
    im = tensor.matrix('im', dtype='float32')
    s = tensor.matrix('s', dtype='float32')
    cim = tensor.matrix('cim', dtype='float32')
    cs = tensor.matrix('cs', dtype='float32')

    # image embedding
    lim = get_layer('ff')[1](tparams, im, options, prefix='ff_im', activ='linear')
    lcim = get_layer('ff')[1](tparams, cim, options, prefix='ff_im', activ='linear')

    # sentence embedding
    ls = get_layer('ff')[1](tparams, s, options, prefix='ff_s', activ='linear')
    lcs = get_layer('ff')[1](tparams, cs, options, prefix='ff_s', activ='linear')

    # L2 norm for sentences
    ls = l2norm(ls)
    lcs = l2norm(lcs)

    # Tile by number of contrast terms
    lim = tensor.tile(lim, (options['ncon'], 1))
    ls = tensor.tile(ls, (options['ncon'], 1))

    # pairwise ranking loss
    cost_im = options['margin'] - (lim * ls).sum(axis=1) + (lim * lcs).sum(axis=1)
    cost_im = cost_im * (cost_im > 0.)
    cost_im = cost_im.sum(0)

    cost_s = options['margin'] - (ls * lim).sum(axis=1) + (ls * lcim).sum(axis=1)
    cost_s = cost_s * (cost_s > 0.)
    cost_s = cost_s.sum(0)

    cost = cost_im + cost_s
    return [im, s, cim, cs], cost

# build an encoder