Mark-Sky / KCL

Implement of 'The Devil is in the Few Shots: Iterative Visual Knowledge Completion for Few-shot Learning'
13 stars 2 forks source link

about the best_beta、best_alpha #2

Open XDS0rryMaker opened 2 months ago

XDS0rryMaker commented 2 months ago

为什么best_beta、best_alpha的生成能够用到测试数据的pse_keys?超参数的生成不是应该和测试数据无关吗?

Mark-Sky commented 2 months ago

You can perceive beta and alpha as regular parameters rather than hyperparameters. The experiment only utilizes the features and pseudo labels of the test data, which is a prevalent approach in semi-supervised learning. Alternatively, you can also keep beta and alpha fixed; however, this might slightly compromise the results.

XDS0rryMaker commented 2 months ago

You can perceive beta and alpha as regular parameters rather than hyperparameters. The experiment only utilizes the features and pseudo labels of the test data, which is a prevalent approach in semi-supervised learning. Alternatively, you can also keep beta and alpha fixed; however, this might slightly compromise the results.

thanks for timely reply. it is of great help!Appreciate!