custom_layers.py 文件源码

def binary_crossentropy(logit, y, mask, length_var):
    logit_shp = logit.shape
    # logit_shp[2] == 1
    # n_labels is 1
    # (n_samples, n_timesteps_f, n_labels)
    # softmax, predict label prob
    # (n_samples * n_timesteps_f, n_labels)
    probs = T.nnet.sigmoid(logit.flatten())
    # (n_samples * n_timesteps_f)
    y_flat = y.flatten()
    loss = lasagne.objectives.binary_crossentropy(probs, y_flat)
    # (n_samples, n_timesteps_f)
    loss = loss.reshape((logit_shp[0], logit_shp[1]))
    loss = loss * mask
    loss = T.sum(loss, axis=1) / length_var
    # (n_samples, n_timesteps_f)
    probs = probs.reshape([logit_shp[0], logit_shp[1]])
    return loss, probs