thinkonward/section-seeker-base-16

Model Card for ThinkOnward's SectionSeeker

This model has been pushed to the Hub using the PytorchModelHubMixin integration:

Model Details

Model Description

The section-seeker-base-16 model is designed to address the challenge of few-shot learning, with a particular focus on one-shot learning scenarios. The model employs a Siamese Neural Network architecture, leveraging a pre-trained ResNet-50 backbone for feature extraction. This architecture allows the model to compare reference images and query images to find matching pairs.

This smaller version (section-seeker-base-16) is part of a twin set of models, which may offer more extensive capabilities or higher accuracy. The section-seeker-base-16 model strikes a balance between computational efficiency and effectiveness, making it suitable for various applications where rapid and accurate predictions are required.

• Developed by: Jakub Mizera, Mike McIntire, Ognjen Tanovic and Jesse Pisel of ThinkOnward
• Model type: ViT
• License: Apache 2.0
• Based on: facebook/vit-msn-base

Model Sources

The section-seeker-base-16 model is built upon several open-source components:

1. Siamese Neural Network Architecture: Based on the architecture described in this paper and this paper.
2. Pre-trained ResNet-50 Backbone: Utilizing weights from the HuggingFace Transformers library.
3. Contrastive Loss: The loss function used for training is implemented using PyTorch, following the methodology described in this paper.

The complete source code and training scripts are available on our GitHub repository. Contributions to improve and extend this model are always welcome!

Uses

The section-seeker-base-16 model is designed for seismic reflection data analysis and has several practical applications in geophysics:

1. Seismic Reflection Data Analysis: The model can quickly match new seismic images with known categories, facilitating faster and more accurate interpretations.
2. Geological Fault Detection: It can detect and classify fault lines within seismic reflection data, aiding in understanding subsurface structural complexity.
3. Reservoir Characterization: By classifying different rock types based on their seismic reflection characteristics, the model provides valuable insights for reservoir characterization.
4. Multi-Modal Data Integration: The model facilitates multi-modal data integration by matching seismic reflection patterns with other datasets, offering a more holistic understanding of subsurface structures.

Out of Scope

• Non-Seismic Data: The model has been specifically trained on and optimized for seismic reflection data.
• Real-Time Applications: While it can process images quickly, real-time processing capabilities are beyond its design scope.
• Very Complex Geological Structures: For extremely complex geological formations where manual analysis might still be necessary.

How to Get Started with the Model

After downloading the model architecture from the SectionSeeker Repository you can load the model using:

import torch
from huggingface_hub import snapshot_download

B16_MODEL_REPO_ID = "thinkonward/section-seeker-base-16"
huggingface_hub.snapshot_download(repo_id=B16_MODEL_REPO_ID, repo_type="model", local_dir='./b_16_checkpoint', allow_patterns='*.pth')


# Use the ModelConfig class from the GitHub repository
vitBConfigPretrained = ModelConfig(BACKBONE_MODEL = 'ViT_B_16',
    BACKBONE_MODEL_WEIGHTS = './b_16_checkpoint/ViT_B_16_SEISMIC31K.pth',
    LATENT_SPACE_DIM = 16,
    FC_IN_FEATURES = 768)
model = SiameseNetwork(vitBConfigPretrained)

Check out the tutorial on GitHub for more help getting started

Training Details

Training Data

The data used to train the SectionSeeker Base 16 model were image patches from real seismic volumes. The data was sourced from the following sources:

• Australia Source: National Offshore Petroleum Information Management System. Available at https://www.ga.gov.au/nopims by Geoscience Australia which is © Commonwealth of Australia and is provided under a Creative Commons Attribution 4.0 International License and is subject to the disclaimer of warranties in section 5 of that license.
• Netherlands Source: NAM (2020). Petrel geological model of the Groningen gas field, the Netherlands. Open access through EPOS-NL. Yoda data publication platform Utrecht University. https://doi.org/10.24416/UU01-1QH0MW

Training Dataset Card: patch-the-planet

Citations

BibTex:

Model Card Contact

Please contact [email protected] for questions, comments, or concerns about this model.