The gradient-based early stopping doesn't seem to work very well as gradients can be low to start, and vary greatly over the course of training. A window of 5 (the default) is too low and results in premature stopping. I'd suggest to either increase the default window or remove it entirely.
The gradient-based early stopping doesn't seem to work very well as gradients can be low to start, and vary greatly over the course of training. A window of 5 (the default) is too low and results in premature stopping. I'd suggest to either increase the default window or remove it entirely.