|
|
--- |
|
|
license: mit |
|
|
--- |
|
|
|
|
|
# Full-Volume Inference without Performance Degradation |
|
|
|
|
|
### π Paper Title |
|
|
**Memory-Efficient Full-Volume Inference for Large-Scale 3D Dense Prediction without Performance Degradation** in **Communications Engineering** (2025) by Jintao Li and Xinming Wu |
|
|
|
|
|
This work introduces a scalable inference framework that enables **whole-volume 3D prediction without accuracy loss**, even on extremely large seismic datasets. |
|
|
The approach restructures high-memory operators during inference only (no retraining required), allowing models to process volumes up to `1024Β³` directly on modern GPUs. |
|
|
|
|
|
The method is particularly useful for seismic interpretation tasks such as fault detection, RGT estimation, implicit structural modeling, and geological feature segmentation. |
|
|
|
|
|
|
|
|
Homepage: [https://github.com/JintaoLee-Roger/torchseis](https://github.com/JintaoLee-Roger/torchseis) |
|
|
|
|
|
--- |
|
|
|
|
|
## π Key Features |
|
|
|
|
|
β **Retraining-free**: works with existing pretrained models |
|
|
β **Whole-volume inference**: no ghost boundaries or stitching artifacts |
|
|
β **Memory-efficient**: reduces decoder/stem memory footprint |
|
|
β **Faster runtime**: avoids slow CuDNN kernel fallback |
|
|
β **Operator-level optimization**: convolution, interpolation, and normalization |
|
|
|
|
|
--- |
|
|
|
|
|
## π Basic Usage |
|
|
|
|
|
Below is an example using `FaultSeg3D` under the TorchSeis framework: |
|
|
|
|
|
```python |
|
|
import torch |
|
|
from torchseis import models as zoo |
|
|
|
|
|
# 1. Load model |
|
|
model = zoo.FaultSeg3d() |
|
|
|
|
|
# 2. Load pretrained weights |
|
|
state = torch.load('faultseg3d-2020-70.pth', weights_only=True) |
|
|
model.load_state_dict(state) |
|
|
|
|
|
# 3. Convert to GPU |
|
|
model = model.half().eval().cuda() |
|
|
|
|
|
# 4. Prepare input volume |
|
|
data = torch.from_numpy(f3d[np.newaxis, np.newaxis].copy()).half().cuda() |
|
|
|
|
|
# 5. Full-volume inference (no tiling) |
|
|
with torch.no_grad(): |
|
|
pred = model(data, rank=3).cpu().numpy() |
|
|
``` |
|
|
|
|
|
> `rank=3` means that using strategy 4 in the paper. |
|
|
|
|
|
--- |
|
|
|
|
|
## π§ Model Zoo Compatibility |
|
|
|
|
|
The full-volume inference method is compatible with the following TorchSeis models, which are available in the [torchseis-efficiency-infer](https://huggingface.co/shallowclose/torchseis-efficiency-infer) on **Hugging Face Hub**. Besides, these models can also be accessed via **Baidu Netdisk**: |
|
|
|
|
|
ιθΏη½ηεδΊ«ηζδ»ΆοΌtorchseis-efficiency-infer |
|
|
ιΎζ₯: [https://pan.baidu.com/s/1ygUPYIO0S1AvsU4-IMz4pg?pwd=y7cn](https://pan.baidu.com/s/1ygUPYIO0S1AvsU4-IMz4pg?pwd=y7cn) ζεη : y7cn |
|
|
|
|
|
|
|
|
|
|
|
| Model | Task | Source | |
|
|
| ---------------- | ---------------------------------------------- | ---------------- | |
|
|
| `FaultSeg3d` | Fault segmentation | [Wu, et, al., 2019, Geophysics](https://library.seg.org/doi/abs/10.1190/geo2018-0646.1) | |
|
|
| `FaultSeg3dPlus` | Fault segmentation | [Li, et, al., 2024, Geophysics](https://library.seg.org/doi/abs/10.1190/geo2022-0778.1) | |
|
|
| `FaultSSL` | Fault segmentation | [Dou, et, al., 2024, Geophysics](https://library.seg.org/doi/abs/10.1190/geo2023-0550.1) | |
|
|
| `Bi21RGT3d` | Relative geological time (RGT) Estimation | [Bi, et, al., 2021, JGR-SE](https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2021JB021882) | |
|
|
| `DeepISMNet` | Implicit structural modeling | [Bi, et, al., 2022, GMD](https://gmd.copernicus.org/articles/15/6841/2022/) | |
|
|
| `ChannelSeg3d` | Channel segmentation | [Gao, et, al., 2021, Geophysics](https://library.seg.org/doi/abs/10.1190/geo2020-0572.1) | |
|
|
| `Wang25Channel` | Channel segmentation | [Wang, et, al., 2025, ESSD](https://essd.copernicus.org/articles/17/3447/2025) | |
|
|
| `KarstSeg3d` | Paleokarst detection | [Wu, et, al., 2020, JGR-SE](https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2020JB019685) | |
|
|
| ~`GEM`~[^1] | ~Geological Everything Model~ | [Dou, et, al. 2025](https://arxiv.org/abs/2507.00419) | |
|
|
| ~`SegFormer3D`~[^2] | ~3D medical segmentation (Transformer)~ | [Perera, et, al., 2025, CVPR](https://arxiv.org/abs/2404.10156) | |
|
|
|
|
|
|
|
|
--- |
|
|
|
|
|
[^1]: **`GEM` Model Note:** The paper is currently under peer review. We will await the official release of the model weights by the authors before considering distribution. |
|
|
|
|
|
[^2]: **`SegFormer3D` Model Note:** As the code variables were not changed, users are advised to use the original author's weights directly. We are uncertain about the rights to redistribute these weight files. |
|
|
|
|
|
--- |
|
|
|
|
|
## π Results Summary |
|
|
|
|
|
The performance evaluations, visual comparisons, and numerical errors reported in our Communications Engineering (2025) paper are fully reproducible. |
|
|
|
|
|
All plotting scripts and evaluation utilities can be found under `scripts/infer25ce/` in the repository: [torchseisJintaoLee-Roger/torchseis](https://github.com/JintaoLee-Roger/torchseis). |
|
|
|
|
|
This directory contains: |
|
|
- **Fig. 3** β Runtime and memory usage comparisons across models |
|
|
- **Fig. S1** β Runtime and memory usage comparisons across models |
|
|
- **Table S1** β Numerical error statistics (full vs. chunked inference) |
|
|
|
|
|
The raw files used to compose Figure 5 is on the zenodo: [https://doi.org/10.5281/zenodo.17810071](https://doi.org/10.5281/zenodo.17810071) |
|
|
|
|
|
--- |
|
|
|
|
|
## π Citation |
|
|
|
|
|
If this work or the inference strategy is used in your research, please cite: |
|
|
|
|
|
```bibtex |
|
|
@article{li2025infer, |
|
|
title={Memory-Efficient Full-Volume Inference for Large-Scale 3D Dense Prediction without Performance Degradation}, |
|
|
author={Li, Jintao and Wu, Xinming}, |
|
|
journal={Communications Engineering}, |
|
|
year={2025} |
|
|
} |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
For questions or contributions, please submit an issue or pull request via the main repository. |
