Training Detail

[SmallYour]

Hi, I am trying to reproduce the baseline(Topdown in Flickr 30k, with CIDEr optimization in RL phase) in your paper, and I used different kinds of code including yours, but I failed, The result of CIDEr score is around 60, can not reach the similar result in your paper (over 67), can you share the detailed training setting?

Thanks!

[YuanEZhou]

Strange! I will check the training setting later and give you feedback within this week.

[YuanEZhou]

Sorry for the later reply. I am running the code and the CIDEr score is around 64 after 60 epochs and I think it will reach 67 after 110 epochs.

To reproduce the baseline (Topdown in Flickr 30k, with CIDEr optimization in RL phase), please try:
python train.py --id CE --caption_model topdown --input_json data/flickrtalk.json --input_fc_dir data/flickrbu/flickrbu_fc --input_att_dir data/flickrbu/flickrbu_att --input_box_dir data/flickrbu/flickrbu_box --input_label_h5 data/flickrtalk_label.h5 --batch_size 29 --learning_rate 5e-4 --learning_rate_decay_start 0 --scheduled_sampling_start 0 --checkpoint_path log/CE --save_checkpoint_every 1000 --val_images_use -1 --max_epochs 30
and then
python train.py --id sc-cider-CE --caption_model topdown --input_json data/flickrtalk.json --input_fc_dir data/flickrbu/flickrbu_fc --input_att_dir data/flickrbu/flickrbu_att --input_box_dir data/flickrbu/flickrbu_box --input_label_h5 data/flickrtalk_label.h5 --batch_size 29 --learning_rate 5e-5 --start_from log/CE --checkpoint_path log/sc-cider-CE --save_checkpoint_every 1000 --language_eval 1 --val_images_use -1 --self_critical_after 30 --max_epochs 110 --cider_reward_weight 1 --ground_reward_weight -1

[SmallYour]

According to your code, we successfully reproduce the performance in your paper. Thanks for replying.