about the 'NAN’ loss during training

[XiejcSYSU]

Hi！why the first version of loss function will cause 'nan' loss during training? Thank you.

[Zhaominghao1314]

'tf.log' resuts in 'inf', just change to 'maxi = tf.log(tf.clip_by_value(tf.nn.sigmoid(pos_scores - neg_scores),1e-8, 1))' in L284 of NDCG.py