Skip to content

Latest commit

 

History

History
48 lines (28 loc) · 1.97 KB

README.md

File metadata and controls

48 lines (28 loc) · 1.97 KB

BoFiCap

NLPCC2023 Best Paper: Bounding and Filling: A Fast and Flexible Framework for Image Captioning

Environment

  1. We use conda to build virtual environment, and we export the env configs to env.yaml, you can reproduce the project env by this yaml.
  2. Besides virtual env, evaluation relies on some metric project, so we build our project based on self-critical.pytorch, you can refer to this project to install metric project, e.g., cider.

Data

  1. We use MSCOCO Dataset as our dataset, and follow the standard Karpathy Splits.
  2. To reproduce our training, you need to preprocess sentence in data into phrase level, and the preprocessed data had been uploaded in data folder.
  3. We also provide preprocess scripts to build phrase datasets, the detailed usage will be upload soon.
  4. AT LAST, your data folder should contain three files:
    1. cocotalk_stanza_kd100_syn_dep0.json
    2. cocotalk_stanza_kd100_syn_dep0_label.h5
    3. cocobu_att.lmdb (detailed info refer to here.)

Training

Our training process including two stages: Cross Entropy Training and Self Critical Training

  1. Cross Entroy Training

    python tools/train.py --cfg configs/uic_sd.yaml --id any_thing_you_like

  2. Self Critical Training

    python tools/train.py --cfg configs/uic_sd_kd100_sd_nscl.yaml --id any_thing_you_like

​ remember to edit the checkpoint path while you using self-critical training

Test

python tools/eval.py --input_json data/cocotalk_stanza_kd100_syn_dep0.json --input_att_dir data/cocobu_att.lmdb --input_label_h5 data/cocotalk_stanza_kd100_syn_dep0_label.h5 --num_images -1 --model model.pth --infos_path infos.pkl --language_eval 0