Wrong trigger for warning "UserWarning: The passed formatting_func has more than one argument."

[zabealbe]

In trl/trainer/utils.py the code func.__code__.co_varnames is being used to check if the user passed a formatting_func with more than 1 parameter. This code actually counts the function variables rather than function parameters.

For instance

def add_v1(a):
  return a + 1

def add_v2(a):
  b = a + 1
  return b

print("Number of parameters for add_v1 is", len(add_v1.__code__.co_varnames))
print("Number of parameters for add_v2 is", len(add_v2.__code__.co_varnames))

outputs

Number of parameters for add_v1 is 1
Number of parameters for add_v2 is 2

The proposed fix is to change the following:

formatting_func_signature = formatting_func.__code__.co_varnames
if len(formatting_func_signature) > 1:
    warnings.warn(
        "The passed formatting_func has more than one argument. Usually that function should have a single argument `example`"
        " which corresponds to the dictionary returned by each element of the dataset. Make sure you know what you are doing."
    )

to:

if formatting_func.__code__.co_argcount > 1:
    warnings.warn(
        "The passed formatting_func has more than one argument. Usually that function should have a single argument `example`"
        " which corresponds to the dictionary returned by each element of the dataset. Make sure you know what you are doing."
    )

Tested on python Python 2.7.5 and Python 3.6.8