Pytorch 在Meta Learning计算中的不足

正如 higher 首页所说

higher is a library providing support for higher-order optimization, e.g. through unrolled first-order optimization loops, of “meta” aspects of these loops. It provides tools for turning existing torch.nn.Module instances “stateless”, meaning that changes to the parameters thereof can be tracked, and gradient with regard to intermediate parameters can be taken. It also provides a suite of differentiable optimizers, to facilitate the implementation of various meta-learning approaches.

即原版的torch.nn.Module不支持对parameters的梯度操作，导致形如像learning to reweight这一类方法用不起来

facebook自己的工程师在也有说明：

Another issue you’re gonna face is that nn.Parameter() are built explicitly to be leaf Tensors (with no history) and so you won’t be able to change them with a Tensor you try to backprop through.

You can check the higher 2 package to do this properly (they solve this problem of nn.Parameter for you). Otherwise, if you want to do it manually, you can find some info in this thread: How does one have the parameters of a model NOT BE LEAFS?

dirty、不用框架的方法

参考：

  1. How does one have the parameters of a model NOT BE LEAFS? 详细的讨论，推荐仔细阅读

不推荐！！

推荐方法：直接用higher框架

参考：

  1. Learning-to-Reweight-Examples-for-Robust-Deep-Learning-with-PyTorch-Higher 直接参考其中的notebook！！推荐！！！