dcn添加偏置，出现Nan错误

[lijun900302]

_**paddle.fluid.core.EnforceNotMet:**_ Invoke operator auc error.

Python Callstacks:
File "/usr/lib64/python2.7/site-packages/paddle/fluid/framework.py", line 1712, in append_op
attrs=kwargs.get("attrs", None))
File "/usr/lib64/python2.7/site-packages/paddle/fluid/layer_helper.py", line 43, in append_op
return self.main_program.current_block().append_op(*args, **kwargs)
File "/usr/lib64/python2.7/site-packages/paddle/fluid/layers/metric_op.py", line 172, in auc
"StatNegOut": [batch_stat_neg]
File "dcn1.py", line 221, in ctr_wide_deep_model
auc_var, batch_auc_var, auc_states = fluid.layers.auc(input=predict, label=label)
File "dcn1.py", line 238, in train
datas, predict, loss, auc, auc_batch = ctr_wide_deep_model()
File "dcn1.py", line 369, in
train()
C++ Callstacks:
Enforce failed. Expected predict_data <= 1, but received predict_data:-nan > 1:1.
The predict data must less or equal 1. at [/home/amy/eason/install/Paddle/paddle/fluid/operators/metrics/auc_op.h:80]
PaddlePaddle Call Stacks:

[wanghaoshuang]

建议您通过在网络中插入Print op定位出现异常的具体某一层。另外，出现NAN也有可能是脏数据导致的。

[mmglove]

请问这个代码是哪个库下面的，有链接吗？

[lijun900302]

请问这个代码是哪个库下面的，有链接吗？
写的dcn, 这里面出现Nan，换个优化算法，又可以跑了
def cross_network(x0,slot_num):
xl=x0
for l in range(cross_layers):
x0_1 = paddle.fluid.layers.unsqueeze(x0, [2], name=None)
xl_1 = paddle.fluid.layers.unsqueeze(xl, [2], name=None)
w = paddle.fluid.layers.create_parameter(shape=[slot_num*(g_embedding_size),1],
dtype='float32', name="w")
xw = paddle.fluid.layers.matmul(xl_1,w,transpose_x=True)
xl_xw = paddle.fluid.layers.matmul(x0_1, xw)
xl_1 = fluid.layers.reshape(xl_xw, shape=[-1, slot_num*(g_embedding_size)], name='flatten1')
xl = xl_1 + xl
bias = paddle.fluid.layers.create_parameter([slot_num*(g_embedding_size)], dtype='float32', name="bias")
xl = xl+bias