YanjieGao / Static-Analyser_ML

Design choices for Shapeflow:

1. Input of our_Operations class should only be lists, not any other objects to make it iterable
2. Input to the forward function of any operation should be the variable (changing) parameters of that operation. The constant ones are not.
3. The asserts involving the variable (changing) parameters (something like x or input_tensor) of the operation is repeated in its main body and its forward function. Other asserts are only in the main body of the operation (something like axis or strides). The main body ones, can catch errors before the session is called and the forward functions ones can do it after.
4. The output of our_Operation is always a Tensor object
5. Placeholders, variables, constants, operations whenever declared are added to the default graph (object of class our_Graph). This is later used when session is called. All operations on objects of class our_Operation (eg. add, sub, getitem) must return an object from same class, so that it can traced back to needed placeholders.
6. If the input to a function (API) is just tensors, then we can return tensors, no need to create an Operation object as it is only needed when the input shape is not known. If input is Operation object then we should return Operation objects.

Suggestions

1. Write the script that can get the tool working on different version of TF. The script should be similar to the one shipped by TF to translate TF 1.0 version code to TF2.0 version code