BatchNorms of some models missing Trainable flag

Some of the models don't use the "trainable" argument in the BatchNorm layers.
While the impact on finetuning models should be neglectable, it should result in lower performance when training them from scratch.