graphsage训练时loss为nan

你好 分布式训练graphsage时 loss一直为nan，mrr不变，相同的数据训练line没问题，参数如下，调小lr到0.001也一样，请问是什么原因呢？
-aggregator='mean'
-batch_size=512
-dim=64
-feature_dim=3
-feature_idx=0
-learning_rate=0.01
-max_id=60000000
-mode='train'
-model='graphsage'
-num_epochs=1
-optimizer='sgd'
-share_negs=False
-unsupervised_loss='rank'


No source file or test is named. Start by reproducing distributed GraphSAGE training with the listed parameters and compare its loss and MRR with the working LINE run; done means identifying why the loss becomes NaN or documenting the missing input or configuration condition.

Written by the indexing model from the issue text.