Compare model weights and visualize differences
Upload a machine learning model to Hugging Face Hub
Teach, test, evaluate language models with MTEB Arena
Submit deepfake detection models for evaluation
GIFT-Eval: A Benchmark for General Time Series Forecasting
Browse and filter machine learning models by category and modality
Merge machine learning models using a YAML configuration file
Evaluate adversarial robustness using generative models
Evaluate reward models for math reasoning
Convert PyTorch models to waifu2x-ios format
Create and upload a Hugging Face model card
Track, rank and evaluate open LLMs and chatbots
Browse and submit LLM evaluations
Vis Diff is a model benchmarking tool designed to help users compare and analyze the differences between model weights. It provides a visual and quantitative understanding of how models differ, enabling developers to identify discrepancies, track changes, and optimize performance.
• Model Weight Comparison: Directly compare the weight parameters of two or more models. • Visualization of Differences: Generate visual representations of weight differences to identify patterns and anomalies. • Support for Popular Frameworks: Compatible with widely used deep learning frameworks such as TensorFlow and PyTorch. • Performance Benchmarking: Analyze how model changes impact performance metrics like accuracy or inference time. • Heatmap Displays: Use color-coded heatmaps to highlight areas of significant difference in model weights. • Version Control Integration: Track model weight changes over time and across different versions.
pip install vis-diff
from vis_diff import load_model
model1 = load_model("model_v1.pth")
model2 = load_model("model_v2.pth")
diff = model1.compare(model2)
diff.visualize()
What models can I compare with Vis Diff?
Vis Diff supports models from popular deep learning frameworks like TensorFlow and PyTorch. Ensure your models are in a compatible format (e.g., .pth, .h5).
How does the visualization work?
The visualization uses heatmaps to represent weight differences. Darker colors indicate larger differences between the two models, while lighter colors show minimal differences.
Can I use Vis Diff for real-time comparison?
Yes, Vis Diff can be integrated into your development workflow for real-time comparison. However, for very large models, you may need to adjust settings to optimize performance.