srebuffi/residual_adapters

residual_adapters's Issues

How to use the adapter in Resnet-50?

Thanks!

validation accuracy not match the published results in the paper

Hi,

I have ran the code and found some surprising results. For training from scratch, fine-tuning and parallel adapter training, the validation accuracy for the aircraft data set I get is 38%, 50%, 44%, however the published test accuracy are 57.10%, 60.34%, 64.21%. The results are lower than the published results.
For the UCF101 data set, the training from scratch, fine-tuning and parallel adapter training validation accuracy I get is 72%, 81%, 75%, however the published test accuracy are 43.27%, 51.20%, 50.94%. The results are higher than the published results.
Finally, the validation results I got for the cifar100 data set match the published results.
I simply ran the code in this github and did not make any changes. Although I am comparing the validation accuracy with the test accuracy, however the difference are quite large. Is there something wrong with the published results? Thanks.

Pretrained model gives low validation accuracy

Hi,

I download the pretrained ResNet-26 and use the code of finetuning to test the validation accuracy on ImageNet, however it gives roughly 0% accuracy.

Batch Normalization layers inconsistently placed in regards to paper

According to "Efficient parametrization of multi-domain deep neural networks" the Batch Normalization (BN) layers are not needed for the parallel configuration but the model it appears that you still use task specific BN layers for the parallel configuration:

elif config_task.mode == 'parallel_adapters' and is_proj: self.parallel_conv = nn.ModuleList([conv1x1(in_planes, planes, stride) for i in range(nb_tasks)]) self.bns = nn.ModuleList([nn.BatchNorm2d(planes) for i in range(nb_tasks)])

Additionally in the series configurations, the BN layer is stated to come before the convolution yet that is not the case in the provided code:

if config_task.mode == 'series_adapters' and is_proj: self.bns = nn.ModuleList([nn.Sequential(conv1x1(planes), nn.BatchNorm2d(planes)) for i in range(nb_tasks)])

question about parallel SVD

Hi Sylvestre Rebuffi,
It's inspiring to read your work Efficient parametrization of multi-domain deep neural networks , but I have a problem understanding section 3.4. Cross-domain adapter compression.
As I understand, you used decomposition to compress the number of parameters, and thus shared information among different datasets. But I am confused about how you realize the perspective of sharing parameters to allow target tasks to communicate. Is this part (SVD) also included in your code? Or is there any important information I've missed? It would be very kind of you to reply.