Martin Görner

I don't see any change in the repro Colab. It is still saving all variables as constants in the graph as far as I can tell. And I did test...

Saving lambas in the ;keras format is not going to work. Lambdas can be arbitrary python code. Prefer using [layers.Rescaling](https://keras.io/api/layers/preprocessing_layers/image_preprocessing/rescaling/)

+1 to what [lbortolotti](https://github.com/lbortolotti) said above: use the .keras format to swwitch backends. Exporting from PyTorch to SavedModel has been done elsewhere though: https://github.com/pytorch/xla/blob/r2.1/docs/stablehlo.md#convert-saved-stablehlo-for-serving We might wan to explore that...

Note: it is debatable wether a contrastive trainer should be in KerasCV. A much better match for it [TensorFlow Similarity](https://github.com/tensorflow/similarity) which actually already implements a contrastive trainer in Keras.

The goal is to be consistent with all other preprocessing layers which use explicit width and height parameters (or x and y) and not size=(...,...) tuples ``` keras_cv.layers.RandomCutout(height_factor, width_factor, ...)...