How do I get there from here?

[Pat F]

New to Spyder as of yesterday actually - but it's working much better than either Emacs or VS (standard) or VSCode did for ML development.

I'm trying to extend the oxford102 training set (102 UK flowers) to 103 - adding a new flower - as part of writing an image recognition engine.  Figured out, to some level of approximation, TFRecords so that I can produce new tdfs compatible data files (locally) with 103 class directories, meaning with photos of the new flower added.

Can successfully load ds_train, ds_valid, and ds_test but currently I'm seeing the following error when I run model.fit:

!next
tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.
  (0) Invalid argument:  Feature: file_name (data type: string) is required but could not be found.
     [[{{node ParseSingleExample/ParseExample/ParseExampleV2}}]]
     [[IteratorGetNext]]
  (1) Invalid argument:  Feature: file_name (data type: string) is required but could not be found.
     [[{{node ParseSingleExample/ParseExample/ParseExampleV2}}]]
     [[IteratorGetNext]]
     [[IteratorGetNext/_4]]
0 successful operations.
0 derived errors ignored. [Op:__inference_train_function_220631]

The following is where, inside .fit(), things are crashing:

https://imgur.com/a/8ecK8sN

I'm trying to figure out how to debug down to the level where the error is propagating?  

[yfprojects]

Hello,

Spyder comes with the ipdb python debugger. Maybe that helps debugging your error.

Kind regards

yfprojects

[Pat F]

Yes, I"ve been using ipdb all along.

I was thinking of something more advanced where Spyder would help me determine just where the error messages originated.  Not that any development environment - whether ML or not  - has ever necessarily been good in that regard.  I mean, one thing you do is to go to the source tree (with an interpreted language like python) and search the source for the messages you're seeing - although that's not what I did here.

I solved the particular problem by revising how I was making the TFRecord features (dict of keys if you will) in the first place.  The debugger though was of the usual little help in getting there.