fixing the probability of applying GaussianBlur

[Separius]

Hi,
I've identified a potential issue within the data augmentation pipeline of DINOv2 that appears to deviate from the expected behavior observed in DINO and MoCo-v3. According to the source code for DINO and MoCo-v3, a Gaussian blur is consistently applied to the first global crop. However, in DINOv2, the probability within the GaussianBlur class is erroneously inverted. This is contrary to the documentation and implementation of RandomApply in TorchVision, where p=1 should indeed apply the transformation consistently.

The rectification is straightforward—alter the probability handling to align with the standard expectations. However, this raises two significant questions:

1. Are the pretrained models also affected by this anomaly in data augmentation?
2. If so, what impact might this have on their performance and reliability?