How to get the logits of the model with a text classification pipeline from HuggingFace?

In the Transformers library provided by Hugging Face, you can efficiently perform model inference using the text classification pipeline (pipeline). By default, the text classification pipeline returns the final prediction results of the model, including labels and their corresponding confidence scores. However, if you need to obtain the model's logits (i.e., the raw scores output by the last fully connected layer, which are typically not yet passed through a softmax transformation), you can achieve this by configuring the pipeline parameters. Below, I will detail how to use the Hugging Face Transformers library to extract logits from a text classification model. First, you need to install the Transformers and Torch libraries (if not already installed):

bash
pip install transformers torch

Next, you can implement the code as follows:

python
from transformers import pipeline

# Create a text classification pipeline using the default model and tokenizer (distilbert-base-uncased)
classifier = pipeline("text-classification")

# To obtain logits, set return_all_scores=True during invocation
results = classifier("I love using transformers for NLP tasks!", return_all_scores=True)

print(results)

After setting return_all_scores=True, the classifier returns the logits for each category. These logits represent the output of the model's final linear layer, typically used before the softmax function. This allows you to inspect the raw scores assigned by the model to each label, which is valuable for applications such as multi-label classification or in-depth analysis of model decision-making.

Example Output:

json
[{
    "label": "LABEL_0",
    "score": -2.4385
}, {
    "label": "LABEL_1",
    "score": 2.4450
}]

The above outlines the basic process and code examples for obtaining model logits from Hugging Face's text classification pipeline. You can adjust the model and configuration parameters as needed to accommodate specific application requirements.