Closed Kyfafyd closed 3 years ago
Hi! You can check this line where features are split into n parts (feature-wise adjustment) and then concatenated with pre-trained class-wise weighted mean (class-wise adjustment). https://github.com/yue-zhongqi/ifsl/blob/2fc15a8233c9ec30080b9eaed6e9c10dcc85e695/MAML_MN_FT/methods/meta_toolkits.py#L156 The c=g(x,d) is implemented differently under different adjustment strategies.
Hi! You can check this line where features are split into n parts (feature-wise adjustment) and then concatenated with pre-trained class-wise weighted mean (class-wise adjustment). https://github.com/yue-zhongqi/ifsl/blob/2fc15a8233c9ec30080b9eaed6e9c10dcc85e695/MAML_MN_FT/methods/meta_toolkits.py#L156
The c=g(x,d) is implemented differently under different adjustment strategies.
Thanks very much
Thanks for amazing work and open source code.
However, I have a question: In supplemental materials, for fine-tuning+IFSL, you calculate
c = g(x, d)
, I wonder where is this function in your code? Thanks very much and looking forward to your reply!