Kurumi233

Results 6 comments of Kurumi233

> hello,sir.Extracting the model tar diretory has an error ,do you have the same question? No, I just have an error I metioned above. Are you extracting the tar file...

> thank you,sir.I have solve my question from yours.maybe ,your question can be handled by this: > model = ResNet(ECABottleneck, [3, 4, 6, 3], num_classes=num_classes, k_size=k_size) > if pretrained: >...

这个作者论文里好像有吧,具体计算方法大概是用通道数c除以类别数n,然后看余数 比如512的通道数,200个类 即 512 / 200 = 2 ... 112 于是有112个类别通道数取3,剩余88个类取2

> Hi, thank you for your codes. I have managed to reproduce the results of vgg16 and resnet50 with **dense layers**. However, when I tried to use **linear layer**, the...

确实,我也有相似的结果,仅使用CELoss得到86.1%左右,标准Res50的结果是84.5%,我认为是DenseLayer增加了复杂度。另一方面,在使用pretrained模型时,MCLoss的权重更小,造成的影响更小,可能是调参的问题。MCLoss在trained from scratch的效果是明显的,可视化结果也有一定说服性,虽然涨点不明显,但是他的思想很棒。

- 这个错误看起来是你冻结了模型的参数,比如require_grad=False - 使用resnet可以考虑加一层feature,类似vgg,或者参考论文直接设置模型,本git只是一个参考,你可以直接把loss作用于任何模型,只要输入最后一层feature以及配置好group num