OptimizerSelector:sgd:weight_decay incorrect configuration makes loading fail

[tomaz-suller]

nasbench301/nasbench301/configspace.json

Lines 244 to 251 in 61c670e

	{
	"name": "OptimizerSelector:sgd:learning_rate",
	"type": "normal_float",
	"log": true,
	"mu": 0.025,
	"sigma": 0.05,
	"default": 0.025
	},

When trying to execute nasbench301/example.py, I get the following error:

TypeError: NormalFloatHyperparameter.__init__() missing 2 required positional arguments: 'lower' and 'upper'

After some time in the debugger, I found out the issue is in the OptimizerSelector:sgd:weight_decay parameter, which indeed does not have these required attributes, but rather two other (mu and sigma). Should they just be replaced with lower and upper? If not, how should we address this issue?

I tried installing nasbench301 for use in NASLib but this error also appears when trying to load the API from NASLib (as expected since this library isn't working properly).

[milkmilk511]

I have the same question. How do you solve it?

[tomaz-suller]

@milkmilk511 I forked NASLib, nasbench301 and ConfigSpace to apply the necessary fixes myself. Check out https://github.com/tomaz-suller/nasbench301/tree/fix-naslib. If you want to use NASLib, I also have a fix-naslib branch in my fork of that.

With what concerns these parameters, if you're only using the library to query the benchmark and not to train models from scratch (which was my use case), they don't really matter, so I just added some numbers so the config would load properly. They're overwritten anyway by hardcoded values in nasbench301/api.py (or at least they should be, but I also fixed that)