Loss becomes nan after few batches

[ashley-mv]

Hello, The number of proposals are reduced to zeros after a few iterations, and later the loss becomes nan. VOC dataset is used for training. Has anyone run into the same problem?

[meijuan1102]

Hello, have you solved this problem? I also met the same problem

[QJSQJS]

Hello, I also met the problem+1, :( . Have you solved this problem?

[QJSQJS]

I found a solution! When the batch decreases, the lr should decrease. The following are my parameters:
batch = 8;
subdivisions = 4;
lr = 0.0001;

It works fine!