Missing error information when using ResNet autoencoder

[collinleiber]

The ResNet autoencoder is only working if the input data is of size (3, x32, x32). If it is not, the decoding will crash without giving an informative error.
Therefore, the size of the input_height should be checked and if it is not correct, a informative error message should be given.

[collinleiber]

Some other issues when working with resnet and image training:

• the encode and decode functions (in deep._utils.py) do not work with images (due to the different shape)
• the model should switch to .eval() state when loading a state_dict