You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi @liopeer, I addressed the second item of this issue: multi-channel image support, following the other packages.
What I did:
EdgeCrafterPackage now accepts num_input_channels and builds the ECViT patch embed for that many channels. It no longer restricts the input to 3 channels.
Pretrained RGB weights are adapted to the requested channel count on load, through a load_state_dict pre-hook that reuses the shared helper the DINOv2, DINOv3, and TIMM packages use.
Note on the stack: this change is stacked on #950 (the first item). GitHub does not let a fork pull request use another fork branch as its base. So this PR is opened against main and shows both commits until #950 merges. After #950 is merged I will rebase this branch onto main so it contains only the multi-channel commit.
Activity
lorinczszabolcs commentedon Aug 22, 2026
Hi @liopeer, I addressed the second item of this issue: multi-channel image support, following the other packages.
What I did:
EdgeCrafterPackagenow acceptsnum_input_channelsand builds the ECViT patch embed for that many channels. It no longer restricts the input to 3 channels.load_state_dictpre-hook that reuses the shared helper the DINOv2, DINOv3, and TIMM packages use.Note on the stack: this change is stacked on #950 (the first item). GitHub does not let a fork pull request use another fork branch as its base. So this PR is opened against
mainand shows both commits until #950 merges. After #950 is merged I will rebase this branch ontomainso it contains only the multi-channel commit.Add multi-scale feature support to edgecrafter ECViT backbone
Add multi-channel input support to edgecrafter ECViT backbone
feat(ecvit): add multi-scale feature support (#950)