This is a Detectron2 implementation of "GLIP: Grounded Language-Image Pre-training" (see the original repo), since the original is implemented in the now deprecated maskrcnn_benchmark and only works with old PyTorch versions.
- This is still a quite bare-bones implementation, I focused on getting training and inference to work with the pretrained GLIP-L model, but none of the other options are supported yet.
- Some settings are not used anymore, others from detectron2 might not be used but are hardcoded instead.
- Performance is not 100% the same as the original implementation, but it's quite close in the tests I did.
- DropBlock is not yet implemented in the FPN.
Download the pretrained GLIP model from the original repo.
When in the same directory as the setup.py file, install the package with pip install -e ..
Load the model with the following code (if using a model trained with the original GLIP code):
from glip_detectron2.config import cfg
from glip_detectron2.utils.checkpoint import init_glip_model
cfg = cfg.merge_from_file('config/pretrained/glip_L_finetune.yaml')
model = init_glip_model(cfg)