How hykrow/ai_sismology works
Project Type and Purpose
• This is a multi-module machine-learning system for seismology, consisting of two main subprojects:
• seisLM: a framework for seismic phase picking, model pretraining/fine-tuning, data conversion, and inference.
• SeisCLIP: a framework for seismic source localization and focal-mechanism estimation, with training, zero-shot, and inference capabilities.
Core Components
A. Data Generation & Conversion Layer (emphasize)
1. Raw Data Sources
• HEMEW-3D simulated zip archives (under seisLM/simulated_data)
• STEAD and other waveform repositories (SeisCLIP_content/SeisCLIP/MIFNO/data)
• CSV metadata files for classification, localization, zero-shot tasks
2. Data Conversion & Preprocessing
• seisLM/simulated_data: python scripts to unzip, convert annotations (.zip → CSV/HDF5)
• seisLM/data_pipeline modules:
• augmentations.py (waveform augment)
• collator.py (batch collator)
• foreshock_aftershock_dataset.py (dataset loader)
• pretrain_dataloaders.py
• SeisCLIP_content/SeisCLIP/MIFNO/data: conversion_format_stead.py, conversion_spectrogramme.py, metadata_format_stead.py
3. Metadata & Indexing
• seisLM/seisLM/utils/data_utils.py (path resolution, file indexing)
• SeisCLIP_content/SeisCLIP/Data/*.csv (train/val/test splits)
B. Training & Fine-Tuning Layer
• Pretraining (seisLM/src/pretrain/pretrain_run.py and SeisCLIP_content/SeisCLIP/Pretrain)
• Fine-tuning (seisLM supports SSL finetune; SeisCLIP finetune_seisclip.py)
• Multi-GPU data-parallel support (train_DDP_audio_stead_two_branch_600.py)
C. Model Foundations & Task-Specific Modules
• seisLM/seisLM/model/foundation (wav2vec2 backbones, conv encoder)
• seisLM/seisLM/model/task_specific (phase-pick & foreshock/aftershock heads)
• SeisCLIP_content/SeisCLIP/Event_classification and Zero_shot model_seismic_clip modules
D. Inference & Evaluation Layer
• seisLM/inference (generate_predictions.py, simulated_generate_* scripts)
• seisLM/seisLM/evaluation (pick_eval, analysis)
• SeisCLIP_content/SeisCLIP/test.py, Demo_zeroshot_pred.ipynb
Component Relationships & Data Flow
• From Datasets → DataPipeline → DataLoader → Model Training/Fine-tuning
• From Trained Checkpoints → Inference Scripts → Predictions/Visualizations
• Evaluation modules consume predictions + ground truth → Metrics
Architectural Patterns & Principles
• Modular pipeline separation: data ingestion, preprocessing, training, inference and evaluation are in distinct directories.
• Configurable via JSON configs (seisLM/seisLM/configs) for reproducibility.
• Extensibility: new models or data sources can plug into existing pipeline modules.