gcn训练或测试较大数据集时报shapes of all inputs must match

[zhixiongning]

在用gcn训练ogbn-products数据集时，测试集数据达几百万，即便训练验证成功，测试时也会报tensorflow.python.framework.errors_impl.Invalid ArgumentError: Shapes of all inputs must match:values[0].shape=[512,100] !=values[1].shape=[511,100]，batch_size为512。
当把测试集减少到一万时，测试成功。同样，当训练集数据量为几百万时，训练一段时间后也会报此问题

[acezen]

谢谢你的issue，为了更方便地找到问题，可以贴下你的测试代码吗？以及更完整的log输出？

[zhixiongning]

脚本是参照example修改的：graph-learn/examples/tf/gcn/train_supervised.py，这个例子用的cora数据集，我们把ogbn-products转换成了相同形式，顶点和边均不带权重，分类数和特征数和数据集匹配，训练点240万，验证点2.4万，测试点2.4万
所有训练，测试，验证的batch_size均设置为512，full_graph_mode False，其余训练参数和example保持一致。
在第一个epoch的iteration到100时报错
报错日志大概如下：
File “tensorflow/python/client/session.py”，line1377，in _do_call
return fn（*args）
File “tensorflow/python/client/session.py”，line1360，in _run_fn
return self._call_tf_sessionrun（options， feed_dict，fetch_list，
File “tensorflow/python/client/session.py” line1453 in _call_tf_sessionrun
return tf_session.TF_SessionRun_wrapper(self._session，options，feed_dict，tensorflow.python.framework.errors_impl.InvalidArgumentError: Shapes of all inputs must match:values[0].shape=[512,100] !=values[1].shape=[511,100]
[[{{node Sum/input}}]

Detected at node Sum/input defined at:
File “train_supervised.py”
trainer.train_and_evaluate()
File "graphlearn/python/model/tf/trainer.py"
loss，train_iterator=self.model.build()
File "graphlearn/python/model/tf/gcn/gcn.py"
src_emb=self.encoders['src'].encode(pos_src_ego_tensor)
File "graphlearn/python/model/tf/encoders/ego_graph_encoder.py"
return self._forward(hiddens)
File "graphlearn/python/model/tf/encoders/ego_graph_encoder.py"
h=self._conv_layers[layer_idx].forward(src_vecs,)
File "graphlearn/python/model/tf/layers/gcn_conv.py
update_vecs=tf.reduce_sum([self_vecs，neigh_vecs]，axis=0)
Node：’Sum/input’
Shapes of all inputs must match:values[0].shape=[512,100] !=values[1].shape=[511,100]
[[{{node Sum/input}}]

[zhixiongning]

将训练参数neighs_num不配置成None，限制每跳采样的邻居数，即不会出现该问题。应该是None 导致采样的点太多了。

[yecol]

Thanks for reporting! We will give a reasonable value as default.

[acezen]

@zhixiongning The neighs_num would set a reasonable value when it's None, the commit will release in v0.12.0