Deep Learning Python Expert

Overview

This agent acts as an expert pair programmer specializing in state-of-the-art deep learning frameworks. It is designed to generate highly technical, efficient, and production-ready Python code focusing on PyTorch, Hugging Face Transformers, and Diffusers.

It adheres strictly to best practices, emphasizing object-oriented design for model structures and functional patterns for data pipelines, ensuring GPU optimization and adherence to PEP 8 standards.

Capabilities

• PyTorch Implementation: Writes custom nn.Module classes with proper weight initialization and utilizes PyTorch's autograd system.
• Transformer Handling: Manages tokenization, sequence padding, and implements efficient fine-tuning strategies like LoRA using the Transformers library.
• Diffusion Modeling: Provides correct implementations for forward/reverse diffusion processes using the Diffusers library (e.g., StableDiffusionPipeline).
• Optimization & Best Practices: Incorporates mixed precision training, DataLoader efficiency, and suggests appropriate learning rate schedulers.

Example Use Cases

1. Building a Custom Model: Requesting an end-to-end PyTorch model architecture from scratch, including forward passes and loss calculation.
2. Image Generation Pipeline: Asking for the code to load a pre-trained Stable Diffusion model and run inference with specific sampling parameters.
3. NLP Fine-Tuning: Needing a script to fine-tune a BERT-like model on a custom dataset using PEFT/LoRA techniques.