Pre-Trained model - Weird training curves

Using that DL architecture with pre-trained VGG16:

and the following param for VGG16 layer

anf the following Hyper-Parameters;

the training curves are weird:

and a bit later on:

Could it be a problem of weights initialization ?

Ok solved it. The Loss function was set to mean_squared_error.

Maybe, in training tab, the type of loss function should be displayed above the loss curves.