dssm.py 文件源码

def compute_loss(output, num_samples, num_entries=6, gamma=500.0):
    """Compute the loss of a dataset, given the output of the DSSM.

    Args:
        output (:class:`lasagne.layers.Layer`): the output of the DSSM
        num_samples (int): the number of samples in the dataset
        num_entries (int): the number of compared papers in the DSSM structure
        gamma (float): the coefficient applied in the softmax of the similarities

    Returns:
        theano.tensor.TensorType: the loss of the dataset
    """
    assert (num_entries > 2)
    assert (num_samples > 0)

    # Post-NN operations to compute the loss
    # First, we extract the first output of each bundle
    mask = np.zeros(num_entries * num_samples)
    mask[::num_entries] = 1
    unmask = np.ones(num_entries * num_samples) - mask
    cited = T.extra_ops.compress(mask, output, axis=0)
    odocs = T.extra_ops.compress(unmask, output, axis=0)

    # We duplicate each row 'x' num_entries-1 times
    cited = T.extra_ops.repeat(cited, num_entries-1, axis=0)
    # Then we compute element-wise product of x with each y, for each bundle
    sims = T.sum(cited * odocs, axis=1)

    # We reshape the similarities
    sims = T.reshape(sims, (num_samples, num_entries-1))
    sims = gamma * sims

    # We take the softmax of each row
    probs = T.nnet.softmax(sims)

    # We compute the loss as the sum of element on the first column
    loss_mask = np.zeros(num_entries-1)
    loss_mask[0] = 1
    loss = T.extra_ops.compress(loss_mask, probs, axis=1)

    return -T.log(T.prod(loss))