It is widely acknowledged that large models have the potential to deliver superior performance across a broad range of domains. Despite the remarkable progress made in the field of machine learning systems research, which has enabled the development and exploration of large models, such abilities remain confined to a small group of advanced users and industry leaders, resulting in an implicit technical barrier for the wider community to access and leverage these technologies. In this paper, we introduce PyTorch Fully Sharded Data Parallel (FSDP) as an industry-grade solution for large model training. FSDP has been closely co-designed with several key PyTorch core components including Tensor implementation, dispatcher system, and CUDA memory caching allocator, to provide non-intrusive user experiences and high training efficiency. Additionally, FSDP natively incorporates a range of techniques and settings to optimize resource utilization across a variety of hardware configurations. The experimental results demonstrate that FSDP is capable of achieving comparable performance to Distributed Data Parallel while providing support for significantly larger models with near-linear scalability in terms of TFLOPS.
During an infectious disease pandemic, it is critical to share electronic medical records or models (learned from these records) across regions. Applying one region's data/model to another region often have distribution shift issues that violate the assumptions of traditional machine learning techniques. Transfer learning can be a solution. To explore the potential of deep transfer learning algorithms, we applied two data-based algorithms (domain adversarial neural networks and maximum classifier discrepancy) and model-based transfer learning algorithms to infectious disease detection tasks. We further studied well-defined synthetic scenarios where the data distribution differences between two regions are known. Our experiments show that, in the context of infectious disease classification, transfer learning may be useful when (1) the source and target are similar and the target training data is insufficient and (2) the target training data does not have labels. Model-based transfer learning works well in the first situation, in which case the performance closely matched that of the data-based transfer learning models. Still, further investigation of the domain shift in real world research data to account for the drop in performance is needed.