loss 为 -nan

[jixianruyizhq]

朋友，你好，在使用 fm 模型中，发现训练数据集比较大小，会出现 train loss 为 -nan 的情况。数据量也不大，就30W，后面想用几百万的数据，感觉不能用呀。这是因为多个小于0的值连乘，导致结果为无穷小么。训练日志如下

参考别人的问题，我降低学习率，依然有这样的问题。但是当我把训练集降低的时候，就正常了。

./xlearn_train data-small/vec-small-train.dat -s 1 -v data-small/vec-small-test.dat -s 1 -x acc -r 0.000001

[------------] xLearn uses 24 threads for training task.
[ ACTION ] Read Problem ...
[------------] First check if the text file has been already converted to binary format.
[------------] Binary file (data-small/vec-small-train.dat.bin) NOT found. Convert text file to binary file.
[------------] First check if the text file has been already converted to binary format.
[------------] Binary file (data-small/vec-small-test.dat.bin) NOT found. Convert text file to binary file.
[------------] Number of Feature: 12001
[------------] Time cost for reading problem: 3.75 (sec)
[ ACTION ] Initialize model ...
[------------] Model size: 468.80 KB
[------------] Time cost for model initial: 0.00 (sec)
[ ACTION ] Start to train ...
[------------] Epoch Train log_loss Test log_loss Test Accuracy Time cost (sec)
[ 10% ] 1 -nan -nan 0.749360 0.31
[ 20% ] 2 -nan -nan 0.749360 0.31
[ 30% ] 3 -nan -nan 0.749360 0.31
[ 40% ] 4 -nan -nan 0.749360 0.32
[ 50% ] 5 -nan -nan 0.749360 0.32
[ 60% ] 6 -nan -nan 0.749360 0.33
[ 70% ] 7 -nan -nan 0.749360 0.31
[ 80% ] 8 -nan -nan 0.749360 0.31
[ 90% ] 9 -nan -nan 0.749360 0.30
[ 100% ] 10 -nan -nan 0.749360 0.32
[ ACTION ] Start to save model ...
[------------] Model file: data-small/vec-small-train.dat.model
[------------] Time cost for saving model: 0.00 (sec)
[ ACTION ] Finish training
[ ACTION ] Clear the xLearn environment ...
[------------] Total time cost: 6.90 (sec)