How aifengheshu/plug-play-modules works
Project type and purpose
• This is an open-source library (“Plug-play-modules”) that aggregates a large set of independent neural-network building blocks (attention modules, convolution variants, normalization layers, down/up samplers, backbones, activations, losses, etc.), each implemented in its own Python file or category folder.
• The README describes its goal: share “2025-year plug-and-play code modules” from top conferences (CVPR, ICCV, ICLR, arXiv, etc.) so that practitioners can easily import and swap in new research components without rewriting code.
File structure
• Top-level components:
• Individual .py files (e.g., AxialAttention.py, BAM.py, CondConv.py…)
• Category folders numbered 3–9:
3. Block (功能模块) – dozens of feature-fusion and token-mixer modules
4. Mamba – a subcollection of “ViM” and “SS-Conv-SSM” blocks
5. Normalization – e.g., BCNorm, RepBN, Crossnorm
6. Backbone – lightweight CNN/Transformer backbones
7. Down & Up – pooling or sampling modules
8. Activation – custom activation functions
9. Loss – custom loss functions
• There is a README.md at the root describing the overall library.
• No build scripts or deployment pipelines are present—this is purely a code-library repo.
README for additional insights
• The README list shows 16 example entries, each with paper title, code link, and a link to a blog explanation.
• All modules rely on PyTorch (implied by naming conventions and typical usage), but no explicit requirements.txt is included.
Detailed capture of architectural elements
• Inside each category boundary, show 3–4 representative module names (e.g., DynamicFilter.py, AGCRN.py in Block; RepBN.py in Normalization).
• That each .py file implements a torch.nn.Module subclass.