hello

It`s alse can change https://github.com/google-research/big_transfer/blob/6c83d6459e009fa89d84c1e904611e9b162e6eff/bit_pytorch/models.py#L29 to `v = torch.var(w, dim=[1, 2, 3], keepdim=True, unbiased=False)` `m = torch.mean(w, dim=[1, 2, 3], keepdim=True)`

请问下你的dic.txt文件的空字符是怎么设置的，能共享下你训练的模型吗，我的loss降到40就下不去了，感觉不是很科学

@wangke1003 您好，我最近也在做这个方向，能方便聊下嘛，试了几个demo感觉都是loss降不下去

我觉得有可能是你测试的config没够改 测试时的图像和训练没有同分布

@gh-andre I have tried but it doesn’t work

there something wrong with cudnn,you can try make it without cudnn.it works for me.

I have same question.when I add a layer in c3d1.0 and finetune the network,i think it will be error ,but it he still runs.I think the author's caffemodel doesn't correspond...