Skip to content

v2.2: Split Model Weights

Pre-release
Pre-release

Choose a tag to compare

View all tags
@Shimmer22 Shimmer22 released this 07 Jan 08:03
· 18 commits to main since this release
v2.2
d99cf66

v2.2 Pretrained Split Weights

Pretrained weights for edge-cloud collaborative VLM inference with bottleneck compression.

📦 Files

  • edge_weights.pth (11 MB) - Edge-side weights

    • Components: MobileNetV2 + StridedProjector + Bottleneck.encoder
    • Parameters: 2.87M

  • cloud_weights.pth (486 MB) - Cloud-side weights

    • Components: Bottleneck.decoder + TransformerUpsampler
    • Parameters: 126.63M

🔧 Configuration

  • Bottleneck dimension: 64
  • Transmission size: 3.1 KB (20× compression from 61 KB)
  • Training dataset: COCO val2017

🚀 Usage

See README for download and deployment instructions.

Temporal checkpoint

  • temporal_pair_best.pth — UCF101 temporal-pair checkpoint
  • SHA-256: a810c4fe91e1fca0b8464478a3c75817da2694f29abefa2727c2f17f92d05ae9
Assets 5
Loading