Check that all arguments are valid

See google/python-fire#168 and
https://gist.github.com/trhodeos/5a20b438480c880f7e15f08987bd9c0f

lm/fire_utils.py

@@ -0,0 +1,35 @@
+"""
+Helpers methods for interacting with python fire.
+https://gist.github.com/trhodeos/5a20b438480c880f7e15f08987bd9c0f
+adjusted for keyword only arguments support
+"""
+import functools
+import inspect
+
+
+def only_allow_defined_args(function_to_decorate):
+    """
+    Decorator which only allows arguments defined to be used.
+    Note, we need to specify this, as Fire allows method chaining. This means
+    that extra kwargs are kept around and passed to future methods that are
+    called. We don't need this, and should fail early if this happens.
+    Args:
+    function_to_decorate: Function which to decorate.
+    Returns:
+    Wrapped function.
+    """
+
+    @functools.wraps(function_to_decorate)
+    def _return_wrapped(*args, **kwargs):
+        """Internal wrapper function."""
+        argspec = inspect.getfullargspec(function_to_decorate)
+        valid_names = set(argspec.args + argspec.kwonlyargs)
+        if "self" in valid_names:
+            valid_names.remove("self")
+        for arg_name in kwargs:
+            if arg_name not in valid_names:
+                raise ValueError("Unknown argument seen '%s', expected: [%s]" %
+                                 (arg_name, ", ".join(valid_names)))
+        return function_to_decorate(*args, **kwargs)
+
+    return _return_wrapped

lm/gpt_2_tf/train.py

@@ -14,12 +14,14 @@
 
 from . import model, sample
 from lm.data import END_OF_TEXT
+from lm.fire_utils import only_allow_defined_args
 
 
 def main():
     return fire.Fire(train)
 
 
+@only_allow_defined_args
 def train(
         run_path,
         dataset_path,