How to evaluate with original model feature output? about pyretri HOT 25 CLOSED

You can simply set the aggregators as "GAP" and set the dim_processors as "Identity".

from pyretri.

Thanks. It can work. But why can't set aggregator as none or delete it directly?

from pyretri.

from pyretri.

cool! However, when I plan to search for feature extraction. Something wrong,

It's about config key, could you tell me where to set it?

from pyretri.

from pyretri.

I add a model in the pyretri\models\backbone\backbone_impl\simclr.py and I want to search a good hyper-params. So, I add my model and extractor in the search\search_modules\extract_dict.py like:

from pyretri.

It seems like my model name does not appear when run pyretri.config.get_defaults_cfg.py

from pyretri.

from pyretri.

Do you mean add this in pyretri\models\backbone.__init__.py:

from .backbone_impl.resnet import ResNet
from .backbone_impl.vgg import VGG
from .backbone_impl.reid_baseline import ft_net
from .backbone_impl.simclr import SimCLR
from .backbone_base import BackboneBase


__all__ = [
    'BackboneBase',
    'ResNet', 'VGG',
    'ft_net',
    'SimCLR',
]

I have added this.

from pyretri.

But your name here is "SimCLR" not "simclr".

from pyretri.

I just add it according to ResNet, in resnet.py the class is ResNet and there are resnet50, resnet101 fun et.al. In my simclr.py the class is SimCLR and there is simclr func. So, I think there are other problems.

from pyretri.

from pyretri.

Thank you very much. I have found the problem.

from pyretri.

from pyretri.

So what's the problem? I am curious.

from pyretri.

Yes. You are right. The changes do not in effect before.

from pyretri.

Now, I also have another question. Is this logs right？

from pyretri.

from pyretri.

It likes it skip all aggregators? Is this ok? Can you give some ideas to do image retrieval and search good hyper-parmas methods with N*C?

from pyretri.

from pyretri.

Thanks for your suggestions! : )

And the GAP or other aggregators will not do anything on this N*C features. So, this doesn't matter if I use GAP for it does need one op in aggregation. Is this right?

from pyretri.

from pyretri.

I'm curious if the aggregators don't work. Why I got this? The different aggregator has different performance.

from pyretri.

And every time I run python main/index.py -cfg configs/pattern_simclr.yaml with pca_whiten, the result is different. Is this because of the training phase in PCA?

from pyretri.

from pyretri.