minizon

I've achieved bleu-1 0.791, bleu-4 0.367, meteor 0.278, rouge_l 0.575, cider 1.22 (beam size 1) on Karpathy's 5000 test set using the following parameter setting: self.caption_model = 'topdown' self.rnn_size =...

Hi, have you worked out this problem? I encountered similar issue when training with SCST, the average training reward gradually rise from a negative number to 0, and the CIDEr...

I suggest you to use the code [https://github.com/ruotianluo/bottom-up-attention-ai-challenger](https://github.com/ruotianluo/bottom-up-attention-ai-challenger) to extract the features. It's easy to deploy.

I've encountered the same problem. The kp mask branch gets really hard to converge. Did we miss something important in the config parameters? 

@Superlee506 Thank you for sharing your code. I've realized that I've made a mistake on the horizontal flip augmentation. I did not change the kp left/right labels when I mirrored...