An keras training loss nan or inf

Can i ask for help with the tensorflow or keras question?
i used this repo:
GitHub - yu4u/noise2noise: An unofficial and partial Keras implementation of "Noise2Noise: Learning Image Restoration without Clean Data" which is keras framework(keras=2.1.6 and tensorflow-gpu==1.5.0)
but get the loss of inf or nan from the 2th or 3th step
do anyone have some idea about this?
the training data is from the guidance of the repo, i think it’s not the problem of the data.

If i use cpu, it works ok

I would recommend asking Keras/Tensorflow-specific questions in their discussion board as you would find the experts there :wink:

thank you for reply, i will look back to this if i have time and i turned to another model using torch