ComVEX: Computer Vision EXpo

Hi there! This is a reimplementation library for computer vision models by PyTorch and Einops. Our mission is to bridge papers and codes with consistent and clear implementations.

What are the pros?

1. Consistent Structure \ Every models share similar building objects:

   • xxxBase: A model's base. For checking common input arguments and storing important variables. Sometimes it can also provide specified weight initialization methods or necessary tensor operations, like patching and flattening images in ViT.
   • xxxBackbone: A model's backbone architecture. It includes every needed components to build the model except the classifier.
   • xxxWithLinearClassifier: xxxBackbone plus a projection head as its classifier. Only accept xxxConfig as its argument. Similar to Huggingface. Might provide some variants for differenet objective in the future.
   • xxxConfig: A configuration for all possible coefficients. It also provides model specializations mentioned in the papers.

2. Consistent Namings for papers and across papers \ To make researchers or developers understand implementations as soon as possible, we tightly follow the names of model components from the official papers and be consistent on common namings across papers.

3. Clear Tensor Operations \ We use Einops for almost all tensor operations to unveil the dimensions of tensors, which are usually hidden in the code, and make our implementations explain by themselves.

4. Clear Arguments \ To expose all possible arguments to users but still remain convenience, we categorize building objects into a hierarchical order with 3 levels listed from bottom to top as below:

   • Basic : The paper-proposed and essential objects that mostly inherit directly from nn.Module or other Basic objects, like ViTBase, MultiheadAttention, SpatialGatingUnit, etc.
   • Wrapper: Intermediate objects or wrappers that organize Basic ones, like TransformerEncoderLayer, PerceiverBlock, MLPMixerLayer, xxxWithLinearClassifier, etc.
   • Model: xxxBackbone and xxxConfig.

   Basic and Model objects are the ones crucial for paper-to-code mappings and model usages, so we require their arguments to be fully explicit to users (list all arguments in __init__ methods). And for the sake of convenience, Wrapper objects can use args or **kwargs to pass down necessary arguments. The overall model structures in term of the number of required arguments will look like a hourglass.

5. Semantic Naming \ Excluding some common names like x for the input tensors, ff_dropout for the dropout rate of feed forward networks, and act_func_name for a string of activation function's name supported by PyTorch, all variables, helper functions, and objects should be named meaningfully.

6. Detailed Model Information \ Every models has its own README.md that provides usages, one-by-one argument explanations, and all usable objects and specializations. The official implementations are provided as well if any mentioned in the official paper.

How to install?

pip3 install comvex

How to use?

Please check out the Usage section detailed in models' own README.md.

How to contribute?

Please check out the CONTRIBUTING.md for details.

Notes

• Continuously implementing models, please check them out under the comvex folder for more details and examples folder for some demos.
• Pull requests are welcome!
• From this issue, inheritance doesn't support in torchscript. Therefore, most of our implementations aren't scriptable. But trace seems that doesn't exist this kind of issue and we will use trace as our default and gradually update our code to make ComVEX a trace-supported library.