• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
  • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Wightman Releases NaFlexViT CLIP Test Model

    Timm library creator shares small NaFLexViT CLIP model trained on CC12M with dual jitter methods.

    RW
    1 Source, 25d ago, first seen 25d ago

    TLDR

    Ross Wightman, creator of the timm PyTorch library, posted that he pushed a new test model from recent OpenCLIP trial runs. He states it is the first of its kind, trained with both patch size jitter in the FlexiViT style and sequence length jitter in the NaFlex style. The small NaFLexViT CLIP model was trained on CC12M. Wightman compared it against a version using a fixed patch size with variable sequence length. The post references additional recent work but provides no further details.

    Combined views

    4.2K

    1 Source, first seen 25d ago

    Combined views

    4.2K

    1 Source, first seen 25d ago

    30 likes
    30 likes
    2 comments
    9 saves
    2 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    2 comments
    9 saves
    2 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    1 Source

    @wightmanrI pushed another unique test model from recent OpenCLIP trial runs. I think the first share of its kind, trained with both patch size jitter (FelxiViT) and sequence length jitter (NaFlex style). A small NaFLexViT CLIP model trained on CC12M. Compared against a fixed patch size w/ variable seq len. There was recent work in both timm 1.0.29 and OpenCLIP to provide patch jitter support. timm side was enhanced with performance improvements, prewarm cache of pinv matrices (for FlexiViT resizing). Also updated to work with pre-patchified NaFlex pipelines that feed a flat nn.Linear embed in the model (so pipeline holds 5D tensors to recover the chosen patch size).

    1 Source

    @wightmanrI pushed another unique test model from recent OpenCLIP trial runs. I think the first share of its kind, trained with both patch size jitter (FelxiViT) and sequence length jitter (NaFlex style). A small NaFLexViT CLIP model trained on CC12M. Compared against a fixed patch size w/ variable seq len. There was recent work in both timm 1.0.29 and OpenCLIP to provide patch jitter support. timm side was enhanced with performance improvements, prewarm cache of pinv matrices (for FlexiViT resizing). Also updated to work with pre-patchified NaFlex pipelines that feed a flat nn.Linear embed in the model (so pipeline holds 5D tensors to recover the chosen patch size).