Noticed that RACL relation components were never added to optimizers. So the current status of RACL runs were only training the cnn layers, disregarding all attentions from learning.
This PR updates the initialization of optimizers to now include these shared layers both for IMN architecture (under other_components as "shared") and RACL architecture (under other_components as "relations")
Noticed that RACL relation components were never added to optimizers. So the current status of RACL runs were only training the cnn layers, disregarding all attentions from learning.
This PR updates the initialization of optimizers to now include these shared layers both for IMN architecture (under other_components as "shared") and RACL architecture (under other_components as "relations")