Overview
Therun_workflow.py script provides a unified interface for executing experiment workflows with three distinct modes: training only, benchmarking only, or a full end-to-end pipeline.
Location: scripts/run_workflow.py
Usage
Arguments
--mode
Type: strChoices:
train, benchmark, fullDefault:
full
Selects the workflow execution mode:
train: Runs only the training step using the specified experiment configurationbenchmark: Runs benchmarking and statistical analysis (skips training)full: Executes the complete pipeline: train → benchmark → statistical analysis
--experiment
Type: strDefault:
baseline
Specifies the experiment configuration name or path to a JSON config file. Used only in training mode.
--stats-repeats
Type: intDefault:
5
Number of repetitions for statistical analysis in benchmark mode.
--seed
Type: intDefault:
42
Global random seed forwarded to both benchmark and statistical analysis steps for reproducibility.
Workflow Stages
Train Stage
Executes the model training pipeline:Benchmark Stage
Runs performance benchmarking:Statistical Analysis Stage
Performs statistical analysis over multiple runs:Examples
Run Full Pipeline
Train Only
Benchmark Only
Quick Test Run
Implementation Details
- Working Directory: All commands execute from the repository root
- Error Handling: Uses
subprocess.run(check=True)to halt on any step failure - Python Executable: Automatically uses the same Python interpreter running the script
- Sequential Execution: Steps run in strict order with no parallelization
Related
- verify_environment.py - Environment verification script
- Training Pipeline - Training module documentation