I followed the code provided in the notebook to do my own research. I found that the loaded checkpoint for cifar10 will project image into a 1000 dimension vector, which is not 512 as reported in README. Is this expected or my interpretation is incorrect.
To do linear eval, do we need to append a MLP, which first projects 1000-d vector to 512-d, interpreted as feature, then projects it to 128-d, followed by a 128-to-10 classifier.
I followed the code provided in the notebook to do my own research. I found that the loaded checkpoint for cifar10 will project image into a 1000 dimension vector, which is not 512 as reported in README. Is this expected or my interpretation is incorrect.
To do linear eval, do we need to append a MLP, which first projects 1000-d vector to 512-d, interpreted as feature, then projects it to 128-d, followed by a 128-to-10 classifier.