Sense-GVT / DeCLIP

Supervision Exists Everywhere: A Data Efficient Contrastive Language-Image Pre-training Paradigm
622 stars 31 forks source link

Filip & DeFILIP? :) #5

Open christophschuhmann opened 2 years ago

christophschuhmann commented 2 years ago

Hi, I just realized that you have Filip and DeFilip implementations, very interesting.

Do you already have results How they compare with Clip and DeCLIP? With respect benchmark scores and compute efficiency? :)

May I also ask about the details of the hardware you used for training?

… And - do you have code for multinode training?

SlotherCui commented 2 years ago

We will publish a technical report on arxiv within two weeks and the result and checkpoints will be released in the same time.

The code has already supported multi-node training, but the ddp framework is being debugged.

TheoCoombes commented 2 years ago

Any update as to when FILIP weights may be released?

zlccccc commented 2 years ago

The FILIP and DeFILIP weights are released new.