pytorch-image-models
pytorch-image-models copied to clipboard
huggingface
Add DependencyViT
From-scratch impl of DependencyViT. Official impl not published. Not competitive with sota hierarchical models, advantage over isometric models is lost due to inability to use F.sdpa, but interesting mechanism. Mainly just ViT with a different attn/block, but the cumulative m and topk pruning makes stuff a bit messy. Block exploits nn.Sequential's untyped intermediate states and feeds tuples through for the cumulative m calculation. I messed around a bit with building dependency trees using the dependency masks, but I have no clue what I'm doing. Currently training weights, have a [email protected]% and playing around with a few implementation details and training recipe. Contacted @dingmyu for reference code and weights but no response yet.
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
@fffffgggg54 is this still in progress ?
This model is on hold for me. It requires several additional deviations from the standard ViT impl that I didn't realize were a part of the model. I will probably take a look at it after I get back from vacation. The current impl can be trained to around 73.X% top-1, around 2% below the reported top-1.