hujie-frank / genet Goto Github PK

Hello, Do you have any comparison with BAM / CBAM attention modules? They claim to perform better than SE: https://github.com/Jongchan/attention-module/

a implementation of GENet based on ResNet-50, achieving the same accuracy.

using framework : Mindspore

GENet_Res50

@hujie-frank To my understanding, global convolution means to set the kernel size equals the current feature size, so this module can only be used on images which have the same sizes for training and testing? Is that right?

I noticed that SENet used two 11Conv after global pool.But In the GENet's Parameter-free version,it not use the 11Conv after global pool. And the two module look very similar. So the main difference of SENet and GENet?

Hello, can you please provide PyTorch implementation of the best performing block you have? (Just the code of the relevant block / module would be sufficient, rather than a whole working script). Thank you.

Hello Jie,
Could you please share the hyper-parameters for GENet on CIFAR10/100, I can not reproduce the result based on hyper-parameters mentioned in here.
Currently, the result(original preact-Resnet110 on CIFAR10/100) I reproduced is better than original paper. But when I add SE modules on it and train with same hyper-parameters, the result I obtained is much less than that in your paper.
top1 err of SE-resnet110 on CIFAR10/100 is 5.42% | 25.94%

Hi there,

It seems that you only release the code with global gathering.

Parameter free: all gather layers use global-average-pooling, e.g. this line.

Parameter (depth-wise-conv): you are using kernel-size that is the same as feature spatial size, e.g. this line

Could you release the code for local-wise gathering? (e.g. downsample with a factor of 8/4/2 and then interpolate during exciting) @hujie-frank

There are some interesting ideas in GENet, nice job!
After reading your paper, I notice that the reported accuracy of SENet and Baseline model (i.e., original ResNets) are quite different from SENet paper (~1.% higher). But there is no data augmentation reported in your experiments. Did I miss something?
Without data augmentation, could GENet still outperform ResNets that much? (only use random crop+flip and 90 epochs). THX.