How parameters are determined

[yinhong0819]

I downloaded the model weights provided in the paper and looked at the config inside. I found that there are many different settings for different models, such as 'seq_length', 'awgn_age', 'awgn', 'seq_len_2d'..., and the settings of each model are different. It's different. I would like to ask how the author decided these parameters in the first place, because I couldn't produce such good results using the config provided by the program code. Thank you.

[ipis-mjkim]

Hello, @yinhong0819 .
Sorry for the late reply.
Actually, in this work, we intended to have the various hyperparmeter combinations, including the network architecture, sequence length, etc., and the ultimate goal at the time was to maximize the gain of the heterogeneous cross-model ensemble.
In my personal experience, I've seen a fairly wide range of hyperparameter combinations perform similarly across the board, as this is a problem where the size of the dataset is still in the realm of not being large enough.
Best Regards.