ra-expman 0.2.7

Manage your experiments exactly as you like

ExpMan: Experiment Manager

expman is a lightweight, functional research experiment management library.

Features

• Functional Core: Pure functions for reliable data management

• Universal Tracking: Metrics (Parquet), Config (YAML), Models (PyTorch), Plots (PD/Matplotlib)

• Modern Web Dashboard:

  • Run Comparison: Overlay metrics from multiple runs on interactive Plotly charts.
  • Artifact Browser: Browse and preview all experiment files including auto-generated model architecture graphs.

  

Installation

• Basic Installation (Lightweight, no Torch/Graphviz):

pip install ra-expman

• Full Installation (With PyTorch support for model saving & architecture graphs) (you can also choose to install torch, torchview and graphviz separately using the basic installation):

pip install "ra-expman[torch]"

Development

To develop expman (with frontend watch and backend reload):

expman dev

Quick Start

1. Run an Experiment:

   from expman import Experiment
   import matplotlib.pyplot as plt
   import torch
   import torchvision
   
   # Initialize
   exp = Experiment("my_experiment")
   exp.log_params({"lr": 0.001, "model": "resnet18"})
   
   # Training Loop
   for i in range(100):
       exp.log_metrics({"loss": 0.5 - i*0.001, "accuracy": i*0.01}, step=i)
   
   # Save Artifacts & Auto-Generate Model Graph
   # Saving is non-blocking (runs in background thread)
   model = torchvision.models.resnet18()
   exp.save_model(model, "final.pt", input_size=(1, 3, 224, 224)) 
   

   For a complete example, check examples/test_run.py.

2. Launch Dashboard:

   expman serve ./experiments
   

   Open http://localhost:8000

3. Interactive Analysis: Load a run directly into a Python REPL with metrics and config pre-loaded:

   expman load experiments/my_experiment/runs/run_001