Feat add pytorch gradient + universal tools

[y-prudent]

Main features:

• add gradient_single function for torch models (used in ODIN) in utils.torch_tools.py
• add useful functions in utils.universal_tools.py (abbreviated to ut) that work on both tensorflow and torch frameworks:
  • is_from function (same API as isinstance) to check the framework of a model or tensor ('torch' / 'keras' / 'tensorflow'), tested in test_universal_tools.py
  • gradient_single function
  • common max, softmax, argmax, one_hot, sign functions (to be completed as required)
• make odin.py almost independant from tensorflow framework using universal_tools

[y-prudent]

Ready for reviewing!

Indeed, importing repetedly torch or tf condiserably slows down the execution, that's a very good point. I removed universal_tools.py and integrated all the corresponding functions as methods of the OODModel class (base.py). During the fit call of an OODModel, torch or tensorflow libraries are now globally imported depending on the model framework, determining the behaviour of each function.

[paulnovello]

Putting operations (argmax, sign etc) as methods of OODModel was not satisfying so I tried to put them in tf_tools and torch_tools and lazily import the appropriate one during OODModel.fit method. The imported module is assigned to a class attribute called "op" to enable its use by child classes.

Hesitated between this solution and another that would be to create a class e.g. UniversalOperator initialized with "framework" argument, whose methods are operations with if statement that calls tf syntax when self. framework=="tensorflow" or torch syntax when self. framework=="torch". An object of this class would be instantiated in the OODModel.fit method and assigned as an attribute (similarly to self.op).

What is your opinion about this?