Extracting intermediate features

[indigoviolet]

Hi,

Thanks to rwightman for this excellent project.

I'm trying to use timm to extract features from resnet50 (for example) and use them in an object detection context (say Retinanet). I looked into the receptive field for the extracted features (for a 320x320 image in this case) and found the following:

       (act1) reduction=2: 64,160,160: RF=7
       (layer1) reduction=4: 256,80,70: RF=35
       (layer2) reduction=8: 512,40,40: RF=91
       (layer3) reduction=16: 1024,20,20: RF=267
       (layer4) reduction=32: 2048,10,10: RF=320

So if I wanted to get some intermediate layers between these, for ex. layer1.0 has RF=19 -- how would I go about doing that? Is there a recommended approach?

One way I'm considering is to construct the Resnet instance, then change its feature_info attribute to have what I want, and then construct the FeatureListNet from that. Is that the right way?

More generally, have you considered a way other than out_indices to provide feature selection. It seems to me that some named notation like out_layers=['act1', 'layer1.0', 'layer1', 'layer2', 'layer3'] would be more convenient.

[indigoviolet]

         rn = timm.create_model("resnet50", pretrained=True)
         rn.feature_info = [{'num_chs': 64, 'reduction': 2, 'module': 'act1'},
                            {'num_chs': 256, 'reduction': 4, 'module': 'layer1.0'},
                            {'num_chs': 512, 'reduction': 8, 'module': 'layer2.3'},
                            {'num_chs': 1024, 'reduction': 16, 'module': 'layer3.5'},
                            {'num_chs': 2048, 'reduction': 32, 'module': 'layer4.2'}]
         fn = timm.models.features.FeatureListNet(rn, flatten_sequential=True)