#Pipeline

Pipeline is the official high-level inference API: you pass a task name and a model ID; it wires Preprocessor + model + post-processing. Use it to explore Hub checkpoints. Drop down when you need generate() knobs or a training loop.

#Common tasks

See the full list in the Pipeline docs. In production, always set model=—do not rely on the implicit default.

#Text: classification

from transformers import pipeline

clf = pipeline(
    "text-classification",
    model="distilbert/distilbert-base-uncased-finetuned-sst-2-english",
)
print(clf(["A thoughtful and funny script.", "Waste of time."]))

A list is batched. Each item has label and score.

#Text: generation

from transformers import pipeline

gen = pipeline(
    "text-generation",
    model="distilbert/distilgpt2",
    max_new_tokens=32,
    do_sample=True,
    temperature=0.8,
)
print(gen("The secret of a good tutorial is")[0]["generated_text"])

max_new_tokens caps new tokens so you neither stop too early nor ramble. Greedy vs sampling is in Text Generation.

For Chinese instructions use Qwen/Qwen2.5-0.5B-Instruct with device_map="auto" and dtype="auto". DeepSeek Instruct models on the Hub are similarly CN-friendly—confirm they are ungated before downloading.

#Vision: image classification

from transformers import pipeline

vision = pipeline(
    "image-classification",
    model="google/vit-base-patch16-224",
)
print(vision("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"))

An ImageProcessor (a Preprocessor) normalizes pixels into model tensors.

#Audio: Whisper-tiny

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="openai/whisper-tiny",
)
# asr("path/to/short.wav")

whisper-tiny is small and ungated—good enough to prove a WAV / mic path. For long or multilingual audio, step up to whisper-small+ and watch VRAM.

#What Pipeline does

raw input  →  Preprocessor.encode / preprocess
           →  PreTrainedModel.forward or generate()
           →  post-process (id→token, softmax, timestamps)

That is the same as hand-wiring AutoTokenizer + AutoModel*. Pipeline saves boilerplate; custom padding, beam search, or multi-turn chat templates are cleaner with AutoClasses.

#When to switch tools

• Local chat only: Ollama is simpler.
• Paper metrics, new heads, your own DataLoader: stay on Transformers.
• Tensor shapes and .to(device): PyTorch tutorial.

#Next steps

• Tokenizers and Models
• Text Generation