Pre-training weights loading about triplet-attention HOT 5 CLOSED

Sorry for the trouble. This is because we renamed the module internally after we trained a few models. I just loaded the weights to check, and these these weights should work with this model definition if you name the module pcam in whatever nn.Module you decide to place triplet attention in.

However, there are many factors that could affect the state dict loading, such as the device you are loading the model to, which weights you're loading, and your usage of nn.DataParallel, so it's hard to say what went wrong. If you can share your code in some form, we might be able to help resolve the issue.

from triplet-attention.

excuse me，I have a same problem and can you provide a e-mail that I send my code to you？thanks

from triplet-attention.

and other question，could i put the triplet_attention location where in the resnet50?

from triplet-attention.

excuse me，I have a same problem and can you provide a e-mail that I send my code to you？thanks

Sorry for the late reply. You can reach me at [email protected].

and other question，could i put the triplet_attention location where in the resnet50?

In principle, you can place the triplet attention module after any other pytorch nn.Module that returns a tensor of shape (N, C, H, W). Triplet attention is a dimensionality preserving module, so it plays nice with other layers.

That said, the recommended location for triplet attention module is right before the residual connection in each Bottleneck module. You can refer to our sample resnet implementation.

from triplet-attention.

Hi,

No matter which definitions that I used for triplet-attention definition, I can't load the pertaining weights.

And could you please let me know which the triplet-attention version that matches the Imagenet Pretraining weights?

your problem have solved? I met same problem

from triplet-attention.