RexNet Features

[hasan-nn]

Hello there, I intend to use RexNet as a Feature extractor ( like for segmentation,object detection).
For rexnet_100 , the attribute 'feature_info' contains:

[{'num_chs': 16, 'reduction': 2, 'module': 'features.0'}, {'num_chs': 38, 'reduction': 4, 'module': 'features.2'}, {'num_chs': 61, 'reduction': 8, 'module': 'features.4'}, {'num_chs': 128, 'reduction': 16, 'module': 'features.10'}, {'num_chs': 185, 'reduction': 32, 'module': 'features.15'}]

Usually there are 5 extraction blocks with output_strides(reductions) = [2, 4, 8, 16, 32]
Given the above, I assume they are as follows:

net = rexnet_100()
Block0 = net.stem
Block1 = net.features[0:3]
Block2 = net.features[3:5]
Block3 = net.features[5:11]
Block4 = net.features[11:16]

and the last block in net.features is a rounded channel expansion to 1280 in rexnet_100 for the classifier head.
Is this correct?
Thank you

[rwightman]

@hasan-nn rexnet is a bit of a weird network to use for feature extractor, it has such small/odd numbered bottleneck channel counts. I didn't use the 1280 net.features as one of the feature pyarmid features because it was so much larger than the rest.

[hasan-nn]

Yes I see, thanks for pointing that out