Feature Parity Between Tensorflow & PyTorch: Seeking Your Input

[jamesdolezal]

Hello Slideflow Community,

As we continually evolve Slideflow to support cutting-edge digital pathology AI research, your feedback and insights are invaluable. Slideflow's dual compatibility with Tensorflow and PyTorch has been a cornerstone of our design, ensuring methodological flexibility and seamless integration of new methods, such as SimCLRv2, StyleGAN3, and DinoV2.

However, the landscape is shifting. PyTorch is rapidly becoming the go-to framework in academic research, as evidenced by the growing number of cited PyTorch repositories compared to Tensorflow. This trend impacts our integration efforts, with recent developments predominantly in PyTorch, covering areas like tissue segmentation, cell segmentation, and advanced GANs.

From Papers With Code: https://paperswithcode.com/trends

We stand at a crossroads: should we focus on maintaining Tensorflow feature parity, or prioritize integrating new PyTorch-based methods? Your use cases, needs, and preferences are crucial in guiding this decision.

We'd love to hear from you:

• How are you utilizing Slideflow in your research?
• Which backend do you predominantly use?
• What are your views on prioritizing feature parity vs. feature expansion?

Your experiences, whether you're implementing new methods or just getting started, matter to us. Each contribution, big or small, enriches our journey together to advance digital pathology research.

We appreciate your time and thoughts as we navigate these important decisions together.