about reproduce performance

[wanboyang]

Thanks for opening source code.
The provided pretrained model can perform the results claimed in your paper, but when I fail to training model on flicker dataset to reproduce the results.

[wanboyang]

I use the default setting and features provided in this project. The result on flicker test split is

the SCST training starts about 30000 iterations, it is interesting to find the cider scores drops and always below 0.60

[wanboyang]

should I do something to reproduce the results performed by pretrained models?

[YuanEZhou]

It seems that you just do the first stage traning(cross entropy loss), you should continue to do the second stage traning(self critical loss).

[wanboyang]

I achieved whole training processing stage, the CIDEr always below 0.62

[YuanEZhou]

Strange! Maybe some hyperparameters were not set correctly in the cleaned version. I will check it later.

[YuanEZhou]

Sorry for the late reply. For better performance on CIDEr, you should comment line12 and uncomment line 11.

[wanboyang]

Thanks for replying