A Fortran-Keras Deep Learning Bridge for Scientific Computing

Figure 3

Offline performance-validation mean squared error (MSE) vs. online performance number of steps until crash. (a) All models. (b) By batch normalization usage. (c) By dropout amount. (d) By leaky ReLU coefficient. (e) By learning rate. (f) By number of dense nodes per layer. (g) By number of layers. (h) By total number of model parameters. (i) By optimizer type.

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)