deep-learning-pytorch

This is a concept-and-convention guide that primes an agent to act as a deep learning expert focused on PyTorch, Hugging Face Transformers, Diffusers, and Gradio. Rather than shipping executable code or tools, it provides a curated set of best-practice principles that improve the quality and correctness of the deep learning code an assistant produces, covering model development, training, evaluation, and deployment.

The guidance is organized by topic: key principles (concise technical style, OOP for architectures and functional style for data pipelines, mixed precision, PEP 8); model development with custom nn.Module classes, autograd, weight initialization, and appropriate losses and optimizers; Transformers and LLM work including attention mechanisms, positional encodings, tokenization, and efficient fine-tuning such as LoRA and P-tuning; diffusion models via the Diffusers library and its pipelines; training and evaluation practices like proper data splits, early stopping, learning-rate scheduling, gradient clipping, and NaN/Inf handling; Gradio integration for interactive demos; error handling and debugging including autograd anomaly detection; and performance optimization with DataParallel/DistributedDataParallel, gradient accumulation, and profiling. It also lists a standard dependency set and project conventions such as modular structure, config files, experiment tracking, and version control.

The intended users are ML engineers, researchers, and developers who want an assistant that follows sound deep learning conventions when writing PyTorch, transformer, or diffusion-model code. It is a single-file reference that points back to the official documentation of each library for up-to-date APIs, so it is best read as an opinionated style layer rather than a comprehensive tutorial.