search

YU1ut / MixMatch-pytorch

Code for "MixMatch - A Holistic Approach to Semi-Supervised Learning"
MIT License
633 stars 129 forks source link

A question on lambda_u #33

Closed zhaozhengChen closed 4 years ago

zhaozhengChen commented 4 years ago

Hi,

In the official implementation, the authors ramp up labmda_u in 16 epochs, but in your implementation, lambada_u is gradually increased in the whole training process(1024 epochs).

YU1ut commented 4 years ago

You can change this line to do that. https://github.com/YU1ut/MixMatch-pytorch/blob/master/train.py#L360

But gradually increasing lambada_u in the whole training process seems to lead to better results according to https://github.com/YU1ut/MixMatch-pytorch/issues/9.

zhaozhengChen commented 4 years ago

Thanks.