EcoTorch

EcoTorch is a lightweight, plug-and-play Python package designed to measure and track the ecological and financial impact of training and evaluating PyTorch models.

Key Features

• Seamless Integration: Track training and evaluation sessions using simple Python context managers.
• Hardware Monitoring: Support for NVIDIA GPUs (via NVML) and Apple Silicon (via custom SMC monitoring).
• Global Impact Tracking: Automatically detects location and uses up-to-date carbon intensity and electricity price data.
• Financial Cost Reporting: Estimates the financial cost of your PyTorch workloads in real-time.
• Efficiency Scoring: Provides a specialized score that balances model improvement with environmental cost.

Installation

Install EcoTorch via pip:

pip install ecotorch

Quick Start

You don't need to rewrite your existing code. Just wrap your training or evaluation loops with TrainTracker or EvalTracker.

import torch
from ecotorch import TrainTracker, EvalTracker

model = ...
train_loader = ...
epochs = 10

# Track Training
with TrainTracker(model=model, epochs=epochs, train_dataloader=train_loader) as tracker:
    # Your training loop here
    initial_loss = 2.5
    final_loss = 0.5

# Calculate efficiency score
score = tracker.calculate_efficiency_score(initial_loss=initial_loss, final_loss=final_loss)
print(f"Training Efficiency Score: {score}")

Documentation

For full documentation, including a getting started guide, API reference, and detailed methodology, please see the docs/ directory:

• Introduction
• Getting Started
• API Reference
• Methodology
• Examples
• FAQ

Contributing

We welcome contributions! Please see our Contributing Guidelines and Code of Conduct.

License

EcoTorch is released under the GNU LGPLv3 License.