Transformers Library

Transformers provides common interfaces for tokenization, model loading, inference and training. A tokenizer produces model inputs, the model holds architecture/weights, and configuration describes model settings.

#Pipeline examples

pythonCopy
from transformers import pipeline

classifier = pipeline("sentiment-analysis")
print(classifier("I enjoyed this course."))
generator = pipeline("text-generation", model="gpt2")
print(generator("In this course, we will", max_new_tokens=20))

These examples download and run models; review the model/license, installed versions and hardware before execution.

#Auto classes for more control

pythonCopy
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

checkpoint = "distilbert-base-uncased-finetuned-sst-2-english"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSequenceClassification.from_pretrained(checkpoint)
inputs = tokenizer(
    ["I enjoyed this course.", "I disliked it."],
    padding=True, truncation=True, return_tensors="pt",
)
model.eval()
with torch.no_grad():
    probabilities = torch.softmax(model(**inputs).logits, dim=-1)
print(probabilities)

#Match the task head

Use a suitable causal-language, masked-language, sequence-classification, token-classification or question-answering model class. A base checkpoint and a task-specific head are not interchangeable.

#Save and reload

pythonCopy
tokenizer.save_pretrained("./my_saved_model")
model.save_pretrained("./my_saved_model")
tokenizer = AutoTokenizer.from_pretrained("./my_saved_model")
model = AutoModelForSequenceClassification.from_pretrained("./my_saved_model")

#Training and optimization

Trainer/TrainingArguments can organize training with prepared train/evaluation datasets. Configure epochs, batch sizes and evaluation using the installed version's API; the source's parameter names may differ in newer releases.

Quantization, attention implementations and device mapping depend on compatible models, packages and hardware. Measure memory, latency and quality instead of treating every optimization as an automatic improvement. Framework support also depends on the specific model and library version.