Fix bugs

[suargi]

Thank you for making your code available.

I have found some bugs:

• line 244 in meta_neural_network_architectures.py #24)
• Backup running statistics is incorrect #42
• Hyperparameters specified in configurations are not the same ones used in the original MAML.
• Update code to latest version of pytorch (where to place scheduler and how to properly retrieve the last lr)
• Bug when adding images to the datasets (first image for each label was always duplicated)
• Bug when normalizing the dataset: dataset is datasets_dict["train"] was never true.

It would be interesting to replicate the study and compare again against the original MAML results.

[suargi]

Fixed some more bugs with training/validation.
Some functions had hard-coded training mode, thus when using validation and test datasets, we were still computing gradients and not using statistics properly.

For now, I reverted the configuration files, so I can easily detect whether I break something