search

Maelic / SGG-Benchmark

A New Benchmark for Scene Graph Generation, targeting real-world applications
MIT License
46 stars 5 forks source link

Configs for SQUAT #38

Closed tutuzeraa closed 1 month ago

tutuzeraa commented 1 month ago

Hello! Very interesting work and repo :)

I'm interested in reproducing the SQUAT model. But, it was not clear for me how can i use the SQUAT relation head for training ... It should be MODEL.ROI_RELATION_HEAD.PREDICTOR SquatPredictor, right?

Also, could you please provide the config files that you used for training/evaluating SQUAT? Is possible that I take the configs from the original work and use in the benchmark or that would require major changes?

Thanks for your time

Maelic commented 1 month ago

Hello! I appreciate your interest in my work! Yes you can use the SquatPredictor for training the Squat model. My config for this model is similar to the default one in here. However, depending on which dataset you want to train you may have to adjust the RHO and BETA values. To do so, I would recommend you to follow my hyperparameter tunning method, specifically changing the model_config to the squat_config in the search space here. Hope this helps and happy coding :)

tutuzeraa commented 1 month ago

Thank you!