model specifications not coherent with the MLB paper

The model configuration is not the same as described in the paper. There is a softmax layer missing at the end of the model. The paper concatenates the attention * vision features for all the glimpses and then pass it through a single linear layer. You use non-linearity both times before and after fusion.

No file or test is named; start by locating the model implementation and comparing its configuration and forward path with the MLB paper. Done means the model includes the paper's final softmax, concatenates attention-weighted vision features across glimpses before one linear layer, and matches the paper's nonlinearity placement.

Written by the indexing model from the issue text.