Hugging Face
定义
Hugging Face provides the Transformers library, Hub (models and datasets), and tools for training and deployment. 它是 a central resource for NLP and multimodal models.
它运行在 PyTorch (and TensorFlow/JAX for some models). Use it to load pretrained BERT, GPT, LLMs, and vision models; fine-tune with the same API; and share your own models and datasets on the Hub. Integrates with RAG and agents via community integrations.
工作原理
安装 transformers、datasets，以及可选的 accelerate（用于分布式和混合精度训练）。加ad a pretrained model and tokenizer by name (例如 from_pretrained("bert-base-uncased")); the library downloads from the Hub if needed. Inference: call model(input_ids) or use pipelines (例如 text classification, summarization). Fine-tune: use the Trainer or native PyTorch loops with your dataset; push the resulting model back to the Hub. The Hub hosts model cards, datasets, and spaces (demos). Tokenizers, configs, and model weights are versioned; you can pin versions for reproducibility.
应用场景
Hugging Face is the default for loading, fine-tuning, and sharing NLP and vision models and datasets.
- Loading and fine-tuning pretrained NLP and vision models
- Sharing and discovering models and datasets on the Hub
- Running inference and building pipelines with minimal code