Cytosine
Hello @dkumazaw ! Thanks for your question. I think the test.py is used as a demo code, which only suggests that their method works. The performance of an architecture should...
Because Cifar and ImageNet have different size! For the former, each picture has a size of 32x32, while the latter has 224x224. Since the ImageNet uses cells searched on Cifar10/100,...
> I have tried to search architectures on cifar100, but validation accuracy was so low : 0.59~0.64 > > Also the found architecture have only skip-connection and pooling layers. >...
As far as I know, you are right! The cells are selected based on alpha; operations with the highest alpha value after search will be used to build the cells....
Yes. You may need to 1) change the stem to reduce the size at the very beginning and 2) use smaller batch size.
I've also noticed this point. Probably it is just a random pick. You may try with other initialization such as Xavier etc. Let me know if you try with different...
We tried and gave up:) GL
> EfficientNet is really designed to be used on images of a specific size, but you can just take the model and apply it (probably without any modifications) to images...
Hello, @tobiascz Thanks for this good question. And I am also confused by applying BSN to a customized dataset. The biggest question of me is, which exact layer of the...
> 跪求大神 收敛指点 ：目前 id=5w 人,m=0.35, loss = 0.390078，acc=0.9918 还是比较差 不然其他训练的效果0.998 你好，可否多讲一点你的实现方法？