NeRV++: An Enhanced Implicit Neural Video Representation

Official Pytorch implementation of NeRV++: An Enhanced Implicit Neural Video Representation.

• NeRV++
  • Tags
  • Overall NeRV++ Framework
  • Disclaimer
  • Documentation
  • Requirements
  • Folder Structure
  • Reproducing Experiments
  • Model Pruning
  • Citation
  • License

Tags

Implicit Neural Representation Neural Codec Neural Video Compression Pytorch

Overall NeRV++ Framework

Disclaimer

Please do not hesitate to open an issue to inform of any problem you may find within this repository. Also, you can email me for questions or comments.

Requirements

Python >= 3.6 Pytorch

All packages used in this repository are listed in requirements.txt. To install those, run:

pip install -r requirements.txt

Folder Structure

nerv-plus-plus-main
├── data/                         # Video data dir
├── docs/asset                    # Documentation figures               
├── selective_scan/               # Selective SSM dir
├── models/                       # Backbones dir
│   └── layers.py                 # Layers
│   └── model_best.py             # NeRV++ backbone    
|   └── ...           
├── requirements.txt              # Requirements
├── utils.py                      # Utility functions
├── train.py                      # Training script
└── main.py                       # Main script

Reproducing experiments

Training experiments

The NeRV++ XS experiment on 'big buck bunny' can be reproduced with, NeRV++ {S, M, L} with {9_16_26, 9_16_58, 9_16_112} for fc_hw_dim respectively.

python train.py -e 300 --lower-width 80 --num-blocks 1 --dataset bunny --frame_gap 1 \
        --outf bunny_ab --embed 1.25_40 --stem_dim_num 512_1 --reduction 2 --fc_hw_dim 9_16_8 --expansion 1 \
        --single_res --loss Fusion6 --warmup 0.2 --lr_type cosine --strides 5 2 2 2 2 --conv_type conv \
        -b 1 --lr 0.0005 --norm none --act swish

Evaluation experiments

To evaluate pre-trained model, just add --eval_Only and specify model path with --weight, you can specify model quantization with --quant_bit [bit_lenght], yuo can test decoding speed with --eval_fps, below we preovide sample commends for NeRV-S on bunny dataset

python train.py -e 300 --lower-width 80 --num-blocks 1 --dataset bunny --frame_gap 1 \
        --outf bunny_ab --embed 1.25_40 --stem_dim_num 512_1 --reduction 2 --fc_hw_dim 9_16_8 --expansion 1 \
        --single_res --loss Fusion6 --warmup 0.2 --lr_type cosine --strides 5 2 2 2 2 --conv_type conv \
        -b 1 --lr 0.0005 --norm none --act swish \
        --weight output/nerv_plus/bunny_ab/.../model_latest.pth --eval_only

Decoding: Dump predictions with pre-trained model

To dump predictions with pre-trained model, just add --dump_images besides --eval_Only and specify model path with --weight

python train.py -e 300 --lower-width 80 --num-blocks 1 --dataset bunny --frame_gap 1 \
        --outf bunny_ab --embed 1.25_40 --stem_dim_num 512_1 --reduction 2 --fc_hw_dim 9_16_8 --expansion 1 \
        --single_res --loss Fusion6 --warmup 0.2 --lr_type cosine --strides 5 2 2 2 2 --conv_type conv \
        -b 1 --lr 0.0005 --norm none --act swish \
        --weight output/nerv_plus/bunny_ab/.../model_latest.pth --eval_only --dump_images

Model Pruning

Evaluate the pruned model

Prune a pre-trained model and fine-tune to recover its performance, with --prune_ratio to specify model parameter amount to be pruned, --weight to specify the pre-trained model, --not_resume_epoch to skip loading the pre-trained weights epoch to restart fine-tune

python train.py -e 100 --lower-width 80 --num-blocks 1 --dataset bunny --frame_gap 1 \
    --outf prune_ab --embed 1.25_40 --stem_dim_num 512_1 --reduction 2 --fc_hw_dim 9_16_8 --expansion 1 \
    --single_res --loss Fusion6 --warmup 0. --lr_type cosine --strides 5 2 2 2 2 --conv_type conv \
    -b 1 --lr 0.0005 --norm none --suffix 107 --act swish \
    --weight output/nerv_plus/bunny_ab/.../model_latest.pth --not_resume_epoch --prune_ratio 0.4

Evaluate the pruned and quantized model

To evaluate pruned model, using --weight to specify the pruned model weight, --prune_ratio to initialize the weight_mask for checkpoint loading, eval_only for evaluation mode, --quant_bit to specify quantization bit length, --quant_axis to specify quantization axis

python train.py -e 100 --lower-width 80 --num-blocks 1 --dataset bunny --frame_gap 1 \
    --outf dbg --embed 1.25_40 --stem_dim_num 512_1 --reduction 2 --fc_hw_dim 9_16_8 --expansion 1 \
    --single_res --loss Fusion6 --warmup 0. --lr_type cosine --strides 5 2 2 2 2 --conv_type conv \
    -b 1 --lr 0.0005 --norm none --suffix 107 --act swish \
    --weight output/nerv_plus/prune_ab/.../model_latest.pth --prune_ratio 0.4 --eval_only --quant_bit 8 --quant_axis 1

Distrotion-Compression result

The final bits-per-pixel (bpp) is computed by $$ModelParameter * (1 - ModelSparsity) * QuantBit / PixelNum$$.

Citation

If you use this library for research purposes, please cite:

@INPROCEEDINGS{,
  author={Ghorbel, Ahmed and Hamidouche, Wassim},
  booktitle={}, 
  title={NeRV++: An Enhanced Implicit Neural Video Representation}, 
  year={2025},
  volume={},
  number={},
  pages={},
}

License

This project is licensed under the MIT License. See LICENSE for more details