Extensive reading of the paper [film: visual reasoning with a general condition layer]

FiLM: Visual Reasoning with a General Conditioning Layer

• Summary ： A characteristic level linear adjustment method is proposed , It has a good effect in visual reasoning tasks

• use ： It can be used for feature merging , For example, dealing with multiple input problems of models

• Realization
  
  The feeling is that one of the features is transformed radially , Then I add （ Adding directly feels that some information will be lost , Therefore, in some articles, we find that sometimes people will change to concate）. Intuitively, one of the features is mapped to the same space as the other through the reorganization of the feature , In this space, the two can be added .

• Network in thesis （ be used for QA）
  

• example
  I'm reading an article recently video caption See the use of the above mechanism in the article （feature-wise linear modulation） Merge features
  
  $h_V,h_S$ It is the characteristic of two different modes （ From vision and sensor respectively ）

• summary
  A feature merging method that is easy to use in visual reasoning tasks ：feature-wise linear modulation