checkpoints receipe about deq-flow HOT 1 CLOSED

Hi @nihalgupta84 ,

Thank you for your interest in this work!

To clarify, T, B, and H are the size of the model, i.e., tiny, base, and huge models, respectively.
The chairs, things, sintel, and kitti are training stages of this model. The models were sequentially trained over flyingchairs, flyingthings, sintel (sintel+things+hd1k+kitti), kitti (finally fine-tuned on kitti).
I offered three checkpoints for the C+T pretraining. So you can see the checkpoint name deq-flow-H-things-test-1/2/3. (But they are from the old version. Please use deq-flow-H-things-test-1x/3x.pth instead.)

Typically, the C+T pretraining and S/K zero-shot evaluation suffices to evaluate an idea if you hope to train new models. So you can refer to this script as a starting point.

Please let me know if you need additional help!

Zhengyang

from deq-flow.