How baiiraam/first_break_picking works
This research prototype turns seismic shot gathers into per-trace first-break picks. The documented user workflow is to preprocess HDF5 data into chunks, train a segmentation model, evaluate it against metrics and baselines, then load a checkpoint for inference or export. Separate product capabilities include explainability and multi-dataset batch training. The graph preserves those branches; several implementations are unsampled, so only documented or indexed relationships are connected.
Main entry points: trainer.py Training loop · preprocessing/pipeline.py HDF5 preprocessing · predictor.py Single-shot predictor · batch/pipeline.py Batch training orchestration
Components 25
Data preparation
- Preprocess command
scripts/preprocess.py - HDF5 preprocessing
src/preprocessing/pipeline.py - Shot processing
src/preprocessing/processor.py - Chunking and manifest
src/preprocessing - Chunked data loading
src/data
Model training
- Training command
scripts/train.py - U-Net model zoo
src/models - Training loop
src/training/trainer.py - Segmentation and pick metrics
src/training/metrics.py - Checkpoint management
src/training/utils/checkpoint.py
Evaluation and analysis
- Evaluation command
scripts/evaluate.py - Evaluation runner
src/evaluation/runner.py - STA/LTA baseline
src/baselines/sta_lta.py - Explainability analysis
src/explainability
Inference and deployment
- Single-shot predictor
src/deployment/predictor.py - Input contract
src/deployment/contract.py - ONNX and TorchScript export
src/deployment/exporters
Batch and shared services
- Seismic configuration
src/config.py - Experiment tracking
src/training/utils/tracking.py - Batch training orchestration
src/batch/pipeline.py - Batch configuration
src/batch/config.py - MLflow tracking service
Other components
- Researcher
- Shot gather input
- Per-trace picks
Connections 24
- Researcher → Preprocess command: runs
- Preprocess command → HDF5 preprocessing: starts
- Shot gather input → HDF5 preprocessing: provides HDF5
src/preprocessing/pipeline.py - HDF5 preprocessing → Shot processing: processes shots
src/preprocessing/pipeline.py - HDF5 preprocessing → Chunking and manifest: writes chunks and manifest
src/preprocessing/pipeline.py - Researcher → Training command: runs
- Training command → Training loop: starts training
- Training loop → Segmentation and pick metrics: computes metrics
src/training/trainer.py - Training loop → Checkpoint management: saves best model
src/training/trainer.py - Training loop → Experiment tracking: tracks run
src/training/trainer.py - Checkpoint management → MLflow tracking service: logs artifacts
src/training/utils/checkpoint.py - Experiment tracking → MLflow tracking service: records experiments
- Researcher → Evaluation command: runs
- Evaluation command → Evaluation runner: starts evaluation
- Researcher → STA/LTA baseline: compares baseline
- Researcher → Explainability analysis: requests analysis
- Researcher → Single-shot predictor: loads checkpoint
- Shot gather input → Single-shot predictor: provides shot
src/deployment/predictor.py - Single-shot predictor → Input contract: reads contract
src/deployment/predictor.py - Single-shot predictor → Per-trace picks: returns picks
- Researcher → ONNX and TorchScript export: exports model
- Researcher → Batch training orchestration: runs batch training
- Batch training orchestration → Batch configuration: loads configuration
src/batch/pipeline.py - Batch training orchestration → MLflow tracking service: monitors runs