Wrap input names in PyLayer

[zimka]

Currently DNN graph structure is inaccessible from NeoML/Python - you can get the list of layers, but you can not read how these layers are connected with each other.
With input names getter it is possible to traverse the whole graph starting from sinks and retrieve it's structure.

It could be used in theory to plot the model, similar to torchviz or maybe even in a plugin for netron.

Also fixes incorrect assertTrue which are supposed to be assertEqual

[FedyuninV]

The idea is really good, but I think it's better to provide list of tuples (input_layer_name, output_index). Current solution doesn't provide enough info if model has layers with multiple outputs (e.g. SplitSomeDimension).

[zimka]

I agree that input_output_index is necessary to retreive precise graph structure, but in my opinion the list of tuples (input_layer_name, output_index) is not the best way to handle the problem.

1. Output index of input is needed only in some specific cases, while in most cases there is either no order at all (single output) or the order does not matter (eltwise sum/prod)
2. "Output Index of Input with Index" is quite clumsy concept, which is not obvious from the first glance. It could be better to wrap CDnnLayerLink, which has more intuitive meaning, but this is out of this PR scope

I have added one more method get_input_output_index and tests for it, please check if you are fine with it

[FedyuninV]

1. Agree. My idea now is to add both input_names and inputs.
2. Agree, but I think we can avoid such awkward names like get_input_output_idx.

Here is the idea:

@property
def inputs(self):
   ''' Tuple in which i'th element contains information about the layer connected to the i'th input of `self` (name, output number)
   '''
   return tuple( ... )

@property
def input_names(self):
   ''' Tuple in which i'th element contains the name of the layer, connected to the  i'th inputs of `self`
   '''
   return tuple( layer_name for layer_name, _ in self.inputs )

The things I don't like in current implementation are Why we have input_names as property, but we have to call method with possibility of wrong idx being passed for getting info about output index? Why these methods are so different?

[zimka]

Yep, these questions sound reasonable.
I have replaced .get_input_output_idx in PyLayer interface with input_links property, it returns tuple of tuples. I have kept the internal method, because it actually mimics CBaseLayer interface.

I think .inputs would be unclear, because it could be layer instances for example, IMO input_links describes it more precisely.
Hope one day there will be pybinded CDnnLayerLinks instead of tuples