search

huangzehao / caffe-vdsr

A Caffe-based implementation of very deep convolution network for image super-resolution
MIT License
273 stars 134 forks source link

你好！我想请教一下。 #16

Closed zhangfudiyi closed 7 years ago

zhangfudiyi commented 8 years ago

我用了60多万张图片，每个scale 2/3/4 按等分的。每个hdf5文件中都包含3个scale的训练样本，为什么我无论怎么训练都无法达到作者提供的模型呢！而且测试set5的图片一般都会相差0.3到0.4左右，大神能不能帮我分析一下什么原因么？

huangzehao commented 8 years ago

Hi, I didn't reproduce the multi scale result too. My results are worse about 0.1~0.2 dB. Actually my training solver is not the best setting, this is one of the reasons. And maybe there are some trick in the training of multi scale. You can email the author for help.

zhangfudiyi commented 8 years ago

clip_gradients 有没有测试其他的下降值，比如0.01 0.2 0.5 。我怀疑是不是这个问题。我的电脑训练太慢了 ，不然我想测试一下。我也邮件问了作者，发过去完全没反应，都不理我。