I had often gotten frustrated trying to parse the code at a bunch of different layers of abstraction just to make one change to a model or just see one detail of the implementation. On top of that, matching pre-trained weights to those edited models was pretty hard for me.
Agrippa consists first of a Python package that takes model architectures written in XML syntax and compiles them to ONNX, a general neural network format. That file can then be imported into PyTorch using the package and trained. Afterward, the model can be published to the zoo on the website. Other users can import entire models or parts of a model into their project. On the zoo, the XML makes it really easy to visualize the models down to every operation.
Overall the markup ideally forces people to make their code more local, so that all of the details for any part of the model are readily accessible. The details are also more explicit - things like initialization and weight sharing. All of the parameters are explicit and named as well.
Agrippa consists first of a Python package that takes model architectures written in XML syntax and compiles them to ONNX, a general neural network format. That file can then be imported into PyTorch using the package and trained. Afterward, the model can be published to the zoo on the website. Other users can import entire models or parts of a model into their project. On the zoo, the XML makes it really easy to visualize the models down to every operation.
Overall the markup ideally forces people to make their code more local, so that all of the details for any part of the model are readily accessible. The details are also more explicit - things like initialization and weight sharing. All of the parameters are explicit and named as well.