Test Time Pooling

[alexriedel1]

Hey!
I recently found about Test Time Pooling in this Inference Example https://github.com/rwightman/pytorch-image-models/blob/master/sotabench.py
being defined here https://github.com/rwightman/pytorch-image-models/blob/f8463b8fa9c0490db093b36acfce71fa2363b8c3/timm/models/layers/test_time_pool.py

has anyone some information about this as i cannot find anything on the internet regarding TTP. For example what it exactly does and how the inference is affected by this

Thanks!

[rwightman]

@alexriedel1 this is where I first saw the idea, https://github.com/cypw/DPNs#mean-max-pooling ... it's sort of like a cheap TTA (test-time-aug) but via one pass through the network with larger image size than default (but fixing the global avg pool at original size).

[alexriedel1]

Thanks!