Transfer learning.

A: Comparison of true and predicted long and short axes for the best 2-conv model (blue) and transfer from VGG19 (red). The black dashed line represents identity. B-C-D: distribution of absolute errors on the test set for all four models and for long axes (B), short axes (C) and orientation (D). E: Loss on validation set during the final training of each model as a function of the number of epochs. F: Loss on validation set as a function of time spent training.