How to Use Pretrained Models from Hugging Face in NLP

To use a pretrained model from Hugging Face in NLP, install the transformers library, then load a model and tokenizer with AutoModelForSequenceClassification and AutoTokenizer. Use the tokenizer to prepare text input and the model to get predictions easily.

Syntax

Here is the basic syntax to load and use a pretrained Hugging Face model for NLP tasks:

• from transformers import AutoTokenizer, AutoModelForSequenceClassification: Import classes to load tokenizer and model.
• tokenizer = AutoTokenizer.from_pretrained('model-name'): Load the tokenizer for the model.
• model = AutoModelForSequenceClassification.from_pretrained('model-name'): Load the pretrained model.
• inputs = tokenizer(text, return_tensors='pt'): Convert text to model input tensors.
• outputs = model(**inputs): Get model predictions.

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased-finetuned-sst-2-english')
model = AutoModelForSequenceClassification.from_pretrained('distilbert-base-uncased-finetuned-sst-2-english')

text = "I love using pretrained models!"
inputs = tokenizer(text, return_tensors='pt')
outputs = model(**inputs)

Example

This example shows how to load a pretrained sentiment analysis model from Hugging Face, tokenize a sentence, and get the predicted sentiment scores.

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load tokenizer and model
model_name = 'distilbert-base-uncased-finetuned-sst-2-english'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Input text
text = "I love using pretrained models!"

# Tokenize input
inputs = tokenizer(text, return_tensors='pt')

# Get model outputs
outputs = model(**inputs)

# Convert logits to probabilities
probs = torch.nn.functional.softmax(outputs.logits, dim=-1)

# Print probabilities for each class
print(f"Negative: {probs[0][0].item():.4f}")
print(f"Positive: {probs[0][1].item():.4f}")

Output

Negative: 0.0001
Positive: 0.9999

Common Pitfalls

• Not installing the transformers library before use causes import errors.
• Using a model name that does not exist or is misspelled leads to loading errors.
• Forgetting to tokenize input text properly results in model errors.
• Not using return_tensors='pt' or 'tf' causes input format issues.
• Confusing model types: use AutoModelForSequenceClassification for classification tasks, not AutoModel.

python

from transformers import AutoTokenizer, AutoModel

# Wrong: Using AutoModel instead of AutoModelForSequenceClassification
model_name = 'distilbert-base-uncased-finetuned-sst-2-english'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)  # This will not output classification logits

text = "Test"
inputs = tokenizer(text, return_tensors='pt')
outputs = model(**inputs)
print(outputs.last_hidden_state.shape)  # Outputs embeddings, not classification scores

# Right way:
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained(model_name)
outputs = model(**inputs)
print(outputs.logits.shape)  # Outputs classification logits

Output

torch.Size([1, 8, 768])
torch.Size([1, 2])

Quick Reference

Summary tips for using pretrained Hugging Face models in NLP:

• Always install transformers with pip install transformers.
• Choose the right model for your task (e.g., classification, generation).
• Load both tokenizer and model with the same model name.
• Tokenize input text with return_tensors='pt' for PyTorch or 'tf' for TensorFlow.
• Use model outputs correctly: outputs.logits for classification scores.

Key Takeaways

Install the transformers library and import AutoTokenizer and AutoModelForSequenceClassification.

Load the pretrained model and tokenizer by specifying the model name from Hugging Face.

Tokenize your input text with return_tensors set to 'pt' or 'tf' before passing to the model.

Use model outputs.logits to get prediction scores for classification tasks.

Avoid using the wrong model class or skipping tokenization to prevent errors.