To my best knowledge, we do not have access to model hyper-parameters such as batch size, learning rate, weight decay etc. in the 'input_data' folder. Thus, my understanding is that we are not expected to use these hyper-parameters when coming up with a complexity measure. However, such information might be useful in one way or another. I wonder if it would be possible to provide model hyper-parameters in the 'input_data' folder as well.
Please correct me if these hyper-parameters for each trained model are already available to us. Thank you!Posted by: haozhe-an @ July 28, 2020, 8:42 a.m.
These parameters are indeed not available on purpose. Although we don't doubt they would be very useful, we believe interesting generalization predictors should rely on the properties of the model only, independently of the parameters used for training.
Thanks for your interest in the competition and good luck!