Parser always producing <END> when predicting synonym operation on test questions

[anette123]

Hi team, making it a separate thread from #1

I have been testing the parser separately (which is getting downloaded from http://vcml.csail.mit.edu/data/ckpt/CLEVR_reason.tgz and seems to be pre-trained as it parses well train questions), and while parser.tools.operation.records contains 'synonym' (and therefore should be able to recogninze this operation), it is always predicting '10' (which is < END > operation) when applied to the test questions which all contain the word 'synonym'. For example, the question
'Is sphere a synonym of shiny?'
gets parsed into

# [[{'operation': '<END>', 'argument': 'shiny'},
#   {'operation': '<END>', 'argument': 'sphere'},
#   {'operation': '<END>', 'argument': '<END>'},
#   {'operation': '<END>', 'argument': '<END>'},
#   {'operation': '<END>', 'argument': '<END>'}]. 

Maybe in fact the parser it is not pretrained on this concept?