timm is a very popular python library for Computer Vision models, with an extensive collection of over 1000 model architectures, pre-trained on Imagenet.
This Streamlit application serves as a user-friendly interface for navigating the myriad models available within the timm library.
Try it here online: https://timm-model-explorer.streamlit.app/
requirements:
pip install streamlit plotly streamlit-plotly-events timm==0.9.12
launch with:
streamlit run streamlit_app.py
This is a first app prototype and can be useful to visualize and search the following stuff:
- Plot model's statistics like:
- Top1, Top5 accuracy on Imagenet dataset
- 8 provided datasets based on Imagenet, to better evaluate robustness and out of domain performances
- Number of parameters
- Inference and training performances of the models
- Get for a selected model the following informations:
- model's configuration
- Model Summary, generated by torchinfo
- Basic code to load the model
- More efficient scatter plot
- Keep scatter plot zoom
- Incorporate missed inference and training stats
- Include model's name tag descriptions
- Optimize for responsive website
- Include links to model papers
- Include architecture visualization (e.g. netron)
- Include other metrics
- Include some inference example (e.g. gradcam)