Compare model weights and visualize differences
Convert and upload model files for Stable Diffusion
Browse and evaluate ML tasks in MLIP Arena
Create and manage ML pipelines with ZenML Dashboard
Analyze model errors with interactive pages
Evaluate reward models for math reasoning
Calculate memory usage for LLM models
Export Hugging Face models to ONNX
Determine GPU requirements for large language models
Display and filter leaderboard models
Request model evaluation on COCO val 2017 dataset
View and submit LLM evaluations
Convert Hugging Face models to OpenVINO format
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.