Confusion about the config

[chauminhnguyen]

Hello,
I am trying to reproduce your code and I am confused about these parameters

input_label_h5: data/cocotalk_label.h5
input_fc_dir: data/cocotalk_clip_RN50_fc
input_att_dir: data/cocotalk_clip_RN50_att

Could you please elaborate on these parameters? Like how can I reproduce cocotalk_label.h5, cocotalk_clip_RN50_fc and cocotalk_clip_RN50_att

[j-min]

Hi Nguyễn,

You can obtain such files by running extraction scripts mentioned here: https://github.com/j-min/CLIP-Caption-Reward#ms-coco

python scripts/prepro_labels.py --input_json data/dataset_coco.json --output_json data/cocotalk.json --output_h5 data/cocotalk

python scripts/clip_prepro_feats.py --input_json data/dataset_coco.json --output_dir data/cocotalk --images_root datasets/COCO/images --model_type RN50

python scripts/clipscore_prepro_feats.py --input_json data/dataset_coco.json --output_dir data/cocotalk --images_root datasets/COCO/images

[chauminhnguyen]

Thank you very much

[chauminhnguyen]

One more question, when i run the command to train mle
python tools/train_pl.py --cfg configs/phase2/$RL_ID.yml --id $RL_ID
After the sanity check, the code calculate scores and then end without training. Do you know what's wrong here?