What blocks are selected in the "swinv2_base_window12_192.ms_in22k" model, if out_indeces = (0,1,2,3)?

[vadori]

Hi,

I need some clarification regarding the block selection in the "swinv2_base_window12_192.ms_in22k" model when setting out_indices = (0,1,2,3).

Swin v2 architecture has a series of stages and each stage has a series of blocks. From my understanding, this setup extracts features from the four different stages of the model. It actually seems that the maximum allowed index in out_indices corresponds to the number of stages. However, I’d like to confirm—does this mean that the last block of each stage is selected for feature extraction?

Additionally, is this a general rule, where the last block in a stage is always used for feature extraction? I understand that in ViT architectures, this issue does not arise, as they consist of blocks without distinct stages, allowing for precise control over which block to select.

Thanks for your insights!

[rwightman]

@vadori yes, when out indices are matched to stages (and thus strides, meaning resolution reduction level) it's always the deepest block for that stride. So last block.

[vadori]

@rwightman great, thank you!!