BiSem

This repository contains the complete model, training-initialization, and evaluation weights used by BiSem, a Vision Transformer-based video semantic communication system.

Repository contents

BiSem/
├── README.md
└── checkpoints/
    ├── bisem/                         # Complete checkpoints for inference
    │   ├── bisem_full_fp32.pth
    │   ├── bisem_full_1b.pth
    │   ├── bisem_full_1.58b.pth
    │   ├── bisem_kf2_fp32.pth
    │   ├── bisem_kf4_fp32.pth
    │   ├── bisem_kf4_1b.pth
    │   ├── bisem_kf4_1.58b.pth
    │   └── bisem_kf8_fp32.pth
    ├── pretrained/                    # Encoder/decoder training initialization
    │   ├── sc_enc/
    │   │   └── vit_base_patch16_224.safetensors
    │   └── sc_dec/
    │       └── SD.pth
    ├── tfhub/                         # I3D model used by FVD evaluation
    └── torch/hub/checkpoints/         # AlexNet/Inception metric weights

BiSem inference checkpoints

The files under checkpoints/bisem/ are complete PyTorch state_dict checkpoints.

All released models use 16-frame RGB clips at 224 × 224, ImageNet normalization, a 64-dimensional semantic channel, and an AWGN channel model.

Checkpoint	Temporal mode	Frames transmitted	Weight mode	Activation mode	Size
bisem_full_fp32.pth	Full	16 / 16	FP32	FP32	432.6 MiB
bisem_full_1b.pth	Full	16 / 16	1-bit	INT8	432.6 MiB
bisem_full_1.58b.pth	Full	16 / 16	1.58-bit	INT8	432.6 MiB
bisem_kf2_fp32.pth	Keyframe	2 / 16	FP32	FP32	436.9 MiB
bisem_kf4_fp32.pth	Keyframe	4 / 16	FP32	FP32	436.9 MiB
bisem_kf4_1b.pth	Keyframe	4 / 16	1-bit	INT8	436.9 MiB
bisem_kf4_1.58b.pth	Keyframe	4 / 16	1.58-bit	INT8	436.9 MiB
bisem_kf8_fp32.pth	Keyframe	8 / 16	FP32	FP32	436.9 MiB

Pretrained initialization weights

The files under checkpoints/pretrained/ initialize a new model before training:

File	Purpose	Size
vit_base_patch16_224.safetensors	Initializes the ViT-B/16 semantic encoder	330.2 MiB
SD.pth	Initializes the MAE-style semantic decoder	101.9 MiB

Evaluation weights

The remaining directories contain third-party feature extractors cached locally so evaluation can run without downloading weights at runtime:

• checkpoints/tfhub/: Kinetics-400 I3D TensorFlow Hub module used for Frechet Video Distance (FVD).
• checkpoints/torch/hub/checkpoints/alexnet-owt-7be5be79.pth: AlexNet backbone used by LPIPS.
• checkpoints/torch/hub/checkpoints/weights-inception-2015-12-05-6726825d.pth: Inception weights used by FID.

Download

Download the complete repository into the BiSem project with huggingface_hub:

from huggingface_hub import snapshot_download


snapshot_download(
    repo_id="Ronnie-lhh/BiSem",
    local_dir=".",
)

Integrity

SHA-256 checksums for the BiSem and initialization weights:

d9675f1a96e7d194416342855678c04385ebbc602200eca2fadb950e9bd08faa  checkpoints/bisem/bisem_full_fp32.pth
fd3a60cc18adc8066a3338b32e8840cc8b410194bd5d2f07aeb2a255dac363b0  checkpoints/bisem/bisem_full_1b.pth
e4ce7366fad41a3c95badce08a5020d31ed3090bf7c3ac9ee98bcb77f1243b71  checkpoints/bisem/bisem_full_1.58b.pth
4cace85ce4482701509d22f035509b89ab0cf5130cc0d14e1e4e2daa32a326ee  checkpoints/bisem/bisem_kf2_fp32.pth
0d45c54c622fff50736539d7c5aa9e3da34f20e817ef5b8d2d311d49464aa4c6  checkpoints/bisem/bisem_kf4_fp32.pth
73f596d21552e6b6e57480d3990d1dc1147cf6595293a8b1c948eb2964fc3141  checkpoints/bisem/bisem_kf4_1b.pth
6251ce712dbda0a157cd31d89483612c6e8d1109b467d47696ae102d2c63d083  checkpoints/bisem/bisem_kf4_1.58b.pth
881c5484151c292a1f706c317f91e9df8bd575d49f56320cd31232b6618b2632  checkpoints/bisem/bisem_kf8_fp32.pth
32aa17d6e17b43500f531d5f6dc9bc93e56ed8841b8a75682e1bb295d722405b  checkpoints/pretrained/sc_enc/vit_base_patch16_224.safetensors
91121a867608491c46cd7aec5e53aeb5472f32c280e5e80d142a4186785b8fb3  checkpoints/pretrained/sc_dec/SD.pth