Code and datasets for SEraser

The code we utilized and the corresponding data have been submitted anonymously. To execute, please proceed as per the following steps:

Setup

In a conda env with pytorch / cuda available, run:

pip install -r requirements.txt

Download

For convenience, we pack the subsets of datasets and they are available './data/(e.g., CamelDeer)' (full link). Please change the ROOT_DIR in fine_tune_blip.py. The pretrained VLFMs is available on Lavis (a widely-used repository).

Inference

The provided fine_tune_blip.py can be run on single-gpu automatically and you need A100 GPU for inference:

python3 fine_tune_blip.py \
--root_dir /path/of/datasets \
--dataset_name dataset_model (e.g., 'Waterbirds_clipB32')

Note that you should make sure that all checkpoints and its config are compiled following Lavis.