about prev2out and evaluation score
DongNaeSwellfish opened this issue · 7 comments
Does the prev2out make big difference in evaluation?
Does anyone had tried this and have evaluation score of this
@DongNaeSwellfish No, it does not effect the performance of the model.
Thanks! I was reproducing the show attend and tell in pytorch and couldn't get the evaluation score same as the paper, so wondered if this problem matters much...
@DongNaeSwellfish If you use pretrained resnet152 instead of vggnet, then the performance of the model will be hugely increased. Good luck!
@yunjey Thanks for your code! If I use resnet152, image feature will be [2048], an one-dimension vector. So it will lose spatial information and can I still attention to it?
Because I use [2048]dimensions of resnet, the result get worse...
@clever-curl-curl did you end up using a different CNN architecture to produce better results? If so, please let me know which pre-trained architecture you used and if you were still able to apply attention. Thanks!!
@clever-curl-curl hi can you please share your code for resnet-152 or resnet-50 ?