Function interface and input_names

[billbrod]

I'm not sure if this is worth an issue, but it's something that can make debugging a complicated pipeline a little difficult: if you have forget to add an input for a function node to its input_names at initialization, a function node will still allow you to set it using node.inputs.in1 = 'foo' and this in1 will show up in the node's report in the working directory, even though the node itself, when it runs, will be unable to see this input (and so it will either use a default value, if it's provided, or fail because there aren't enough inputs).

For example:

def func(in1,in2=0):
    return in1+in2

f_node = pe.Node(interface=util.Function(input_names=['in1'],output_names=['sum'],function=func),name='node')
f_node.inputs.in1 = 1
f_node.inputs.in2 = 1

Then, if you try to run this node, it will run without any errors and return a 1, instead of the expected 2. Furthermore, if you look in the report in the node's directory, in2 = 1 will be listed as an input; loading the node using node=loadpkl(_node.pklz) and checking node.inputs will show the same thing. It's only if you run the node and debug it will you notice that in2=0.

Doesn't it make sense to prevent users from being able to add values to node.inputs that aren't listed in input_names? Why does this behavior happen?

This was with nipype version 0.10.0

[djarecka]

Do we consider this as a bug?

[satra]

@billbrod and @djarecka - under current nipype you can skip the input_names argument and all args from the function are used. so the following will return 2.

in the original example, the node only gets one input. the second input doesn't exist as an input. the reason it's accepted is because the dynamic trait is allowing any set of input names.

def func(in1,in2=0):
    return in1+in2

f_node = pe.Node(interface=util.Function(output_names=['sum'],function=func),name='node')
f_node.inputs.in1 = 1
f_node.inputs.in2 = 1

the bug is perhaps in the code that checks if the keyword should be allowed to be assigned. it should be restricted to the options that are input_names.

[mgxd]

@djarecka i remember we sat down and went through this one day long ago - can't find any trace of it though!

[djarecka]

@mgxd : I also remember that we were sitting on something similar but wasn't sure if we decided to leave as it is, or we simply didn't finish. Will try to find it

[effigies]

@satra Another option is to always use the implicit behavior and just raise an node-construction-time error if input_names doesn't match the parameter names. Unless hiding inputs with default values is behavior we want to support.

[satra]

@effigies - i think i wanted this f_node.inputs.in2 = 1 to raise a KeywordError exception rather than fail silently. it's the silent failure that i am worried about it.