Question about shape_list function in modeling_tf_utils

Hi !

I’m trying to use TFBertModel in Rasa framework, and I found some minor mismatch with function ‘shape_list’ in ‘’ and function ‘call’ in TFBertModel

Argument typing says it is okay to pass Union of np.array and tf.Tensor, but I found that when I pass ‘attention_mask’ with np.array, It causes some error in ‘shape_list’ function in ‘’

the error is about np.array instance doesn’t have ‘as_list()’ function

So I changed the line 1958 like this

# static = tensor.shape.as_list()  # original code
   static = list(tensor.shape)        # my implementation

I think it didn’t cause any problem and model worked well

So I was wonder about is there any special reason for using tensor.shape.as_list() which only works for tf.Tensor instance

and I thought It would be nice to change it like I did

function ‘input_processing’ in ‘’ do not make np.array kwarg s into tf.Tensor so I think my implementation would be nice to be merged.

How do you think about it guys?
Of course I might be wrong so feel free to answer my question!

Thank you! :grin:_smile: