Why data used in TCT stage 2 are subsampled?

Hi, thanks for sharing the codes.

We notice that in [utils.py L164](https://github.com/yaodongyu/TCT/blob/e86dfb3e7f28a40b572234fe6a85cbaf846b6d54/mnist/utils.py#L164), only a subset of samples are used for client training. This is not mentioned in the paper. So we are wondering if this is an error in code or some special problem setting?

Many thanks.