Some questions about Advanced Machine learning/Final project_CLIP.ipynb->> Q7: Fine tuning on CIFAR

[MysteriousL2019]

Since in Advanced Machine learning/Final project_CLIP.ipynb->> Q7: Fine tuning on CIFAR part two, code we have

Here the ground_truth is generated from 0 to len(image). And this operation will make the cosine similarity to become the uniform distribution. Why this will enlarge the accuracy?