Deserialization error with Lambda layers when loading saved model in new environment

[bourcierj]

When one saves a converted Keras model with Lambda layers to disk and loads it in a completely separate environment, one will get errors like TypeError: Exception encountered when calling layer "LAYER_173_CHW" (type Lambda). This is because Lambda layers are fundamentally non-portable across environments (cf. Keras docs).

The root cause is given in warnings right before the exception is raised:

[...]UserWarning: onnx2keras.pooling_layers is not loaded, but a Lambda layer uses it. It may cause errors.
To solve this: replace all `Lambda` layers with native or custom layers.
[...]UserWarning: onnx2keras.reshape_layers is not loaded, but a Lambda layer uses it. It may cause errors.

To solve this, we need to replace all Lambda layers with native or custom layers (or at least in this example where the model is ResNet18, those related to pooling and reshape layers).
However Lambda layers are everywhere in this library, so this could require a great amount of work to cover all cases.

Steps to reproduce:

1. convert ResNet18 from PyTorch to Keras (similarly to what's done in test.models.test_resnet18.test_resnet18)
2. save the keras model to a SavedModel format with tf.keras.Model.save(..., save_format="tf")
3. load the saved model with tf.keras.models.load_model in a completely separate environment