conda create --name labert python=3.7
conda activate labert

conda install pytorch=1.3.1 torchvision cudatoolkit=10.1 -c pytorch
pip install h5py tqdm transformers==2.1.1
pip install git+https://github.com/salaniz/pycocoevalcap

python -m torch.distributed.launch \
  --nproc_per_node=$NUM_GPUS \
  --master_port=4396 train.py \
  save_dir $PATH_TO_TRAIN_OUTPUT \
  samples_per_gpu $NUM_SAMPLES_PER_GPU

python -m torch.distributed.launch \
  --nproc_per_node=$NUM_GPUS \
  --master_port=4396 train.py \
  save_dir $PATH_TO_TRAIN_OUTPUT \
  samples_per_gpu $NUM_SAMPLES_PER_GPU \
  model_path $PATH_TO_MODEL

python inference.py \
  model_path $PATH_TO_MODEL \
  save_dir $PATH_TO_TEST_OUTPUT \
  samples_per_gpu $NUM_SAMPLES_PER_GPU

python evaluate.py \
  --gt_caption data/id2captions_test.json \
  --pd_caption $PATH_TO_TEST_OUTPUT/caption_results.json \
  --save_dir $PATH_TO_TEST_OUTPUT

@article{deng2020length,
  title={Length-Controllable Image Captioning},
  author={Deng, Chaorui and Ding, Ning and Tan, Mingkui and Wu, Qi},
  journal={arXiv preprint arXiv:2007.09580},
  year={2020}
}