LossRecord

[LibertFan]

Is the figure of loss really the training loss ? It seems too small. Is your model loss calculated according to the paper ? Or you calculate it in word level instead of paragraph level ?

[Wentong-DST]

Hi LibertFan, The loss is the training loss, and the validation loss is higher than it. So obviously there is an overfitting problem. I think the loss is calculated in word level. The original repo author has ceased any support, so I would recommend you to move onto other image captioning models. Please check Luo Ruotian's implementation of self-critical model.