philmitchell47

3 ms·

1. 1.▲
   by philmitchell47 2y ago

   Seems horrendously brittle. I'd trust badly written bash over this.
2. 2.▲
   by philmitchell47 3y ago

   When they say modern, which standard to they mean?
3. 3.▲
   by philmitchell47 3y ago

   I think it's kind of disingenuous maybe to claim such improvements in training efficiency when they rely on: - Existing models for data pseudo-labelling - ImageNet pretraining - A frozen text encoder - A frozen image encoder