现象:AttributeError: ‘module’ object has no attribute ‘LinearOperatorTriL’
原因:Tensorflow 版本问题
解决方法:
把 tril = tf.contrib.linalg.LinearOperatorTriL(diag_vals).to_dense() # (T_q, T_k)
换成 tril = tf.linalg.LinearOperatorLowerTriangular(diag_vals).to_dense() # (T_q, T_k)
参考: https://github.com/Kyubyong/transformer/issues/30