Is "input" a keyword that causes errors when used as a parameter name (in PyTorch)?

Question

So I have a line of code:

packed_embeddings = pack_padded_sequence(input=embeddings,
                                                 lengths=lengths,
                                                 batch_first=True)

That throws me this error:

  File "/Users/kwj/anaconda3/lib/python3.6/site-packages/torch/onnx/__init__.py", line 130, in might_trace
    first_arg = args[0]

IndexError: tuple index out of range

But magically fixes itself if I take out the "input":

    packed_embeddings = pack_padded_sequence(embeddings,
                                             lengths=lengths,
                                             batch_first=True)

Here is the function specification in the PyTorch docs:

https://pytorch.org/docs/stable/_modules/torch/nn/utils/rnn.html#pack_padded_sequence

I'm using Python3 and PyTorch 0.4. Am I missing something really basic? Not sure if this is my issue, or a PyTorch specific issue...pretty confused here.

Thanks

Answer 1

What's happening here is that pack_padded_sequence is decorated to return a partially applied function, and within the decorating code there is a function that accepts arguments as *args, **kwargs. This function passes args to another function, which inspects the first arg. When you pass all the arguments to packed_padded_sequence as keyword arguments, args is empty, and so args[0] raises an IndexError. If you pass input as a positional argument,args is not empty, and the IndexError is not raised.

This example code demonstrates the behaviour (the Pytorch code is not easy to read).

def decorator(func):
    def wrapper(*args, **kwargs):
        print('Args:', repr(args))
        print('Kwargs:', repr(kwargs))
        return func(*args, **kwargs)
    return wrapper   

@decorator
def f(a, b=0, c=0):
    return a, b, c


if __name__ == '__main__':
    print('Positional argument...')
    print(f(1, b=2, c=3))
    print('All keyword arguments...')
    print(f(a=1, b=2, c=3))

The code produces this output:

Positional argument...
Args: (1,)                        <- Args is populated
Kwargs: {'b': 2, 'c': 3}
(1, 2, 3)
All keyword arguments...
Args: ()                          <- Args is empty
Kwargs: {'a': 1, 'b': 2, 'c': 3}
(1, 2, 3)