I'm having the same issue here. it runs the first training sample then fails the second and subsequent ones.

Running training in RStudio works fine with keras and tensorflow with many different model structures.

The GPU is in use for that first training loop that works so I don't think its a tensorflow/keras issue as such, maybe more an issue saving?

Is there a way to make the logs more verbose in terms of any R logs? 7+DIAG only gives the zorro logs.