search

ruotianluo / self-critical.pytorch

Unofficial pytorch implementation for Self-critical Sequence Training for Image Captioning. and others.
MIT License
993 stars 278 forks source link

regarding scores for uploaded models #80

Open homelifes opened 5 years ago

homelifes commented 5 years ago

Hi. Thanks for your great work! I would like to ask regarding the models uploaded in google drive for att2in2 and top-down. What modifications from the authors paper have you used in both to get these results? Are these implementations as described in the paper or did you add some more features? For att2in2, ive read that you used two MLP layers for the embedding. Apart from that is there anything? and for the top-down model is there any extra feature used in your code that is not reported in the paper? (Please note that I am talking about the pre-trained models you uploaded). Waiting for your kind reply! Thanks!

ruotianluo commented 5 years ago

For att2in2, that's the only change. For top-down, I actually implemented it with my own "guess"; I never check the official repository to see what the discrepancies are. (Features are exactly the same.)

homelifes commented 5 years ago

Thank you for your answer.

homelifes commented 5 years ago

@ruotianluo i also want to ask. 1- Is the top-down uploaded model trained with cider or cross entropy? 2-And did you use bottom up or resent features for the att2in2 and top-down implementation? Thanks!

homelifes commented 5 years ago

@ruotianluo waiting for your kind reply.

ruotianluo commented 5 years ago

@homelifes 1 I think I put them under different folders 2 the pertrained models dont use bottom up I remember.