Implement DeepSeek-R1 Code Analysis and Refactoring

Using DeepSeek-R1, implement a core agent capability to analyze a provided PyTorch training loop and identify at least two common performance bottlenecks (e.g., inefficient data loading, CPU-GPU transfer issues, sub-optimal tensor operations). The agent should then propose concrete code changes to address these bottlenecks, providing a clear explanation for each. Focus on leveraging DeepSeek-R1's strengths in code understanding.