Implement details of MMN w/o BS about meta-module-network HOT 3 OPEN

Are the programs that are shared in this repo also generated by a model that was trained only on train? So, all json files with "fully" in their name (e.g. trainval_unbiased_fully_inputs.json) are generated by a model that was not trained on val set?

from meta-module-network.

Besides, the best testdev accuracy of your provided pre-trained models TreeSparsePost2FullValidationRemoved is model_ep4_0.5623, why this result is not matching to 56.5 of MMN w/o FT ?

from meta-module-network.

The models released in the github is trained on training set only. The paper uses the model trained on the combination of training+validation set, which bumps the score by 2%. "TreeSparsePost2FullValidationRemoved" means the full validation set is removed during training.

from meta-module-network.