Hugging Face: Hub, Transformers & Production Workflows

from transformers import pipeline

# Instantiate high-level task pipeline
classifier = pipeline(
    task="sentiment-analysis",
    model="distilbert/distilbert-base-uncased-finetuned-sst-2-english"
)

# Execute inference on raw string input
result = classifier("Pathubs provides world-class AI engineering education!")
print(result)
# Output: [{'label': 'POSITIVE', 'score': 0.9997}]

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

MODEL_ID = "distilbert/distilbert-base-uncased-finetuned-sst-2-english"

# 1. Load Tokenizer & Model
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)

# 2. Tokenize with explicit PyTorch tensor return
raw_text = "The microservice handles 10,000 requests per second with zero errors."
inputs = tokenizer(
    raw_text,
    return_tensors="pt",       # Return PyTorch tensors (pt)
    padding=True,              # Pad to longest sequence
    truncation=True,           # Truncate at max model length (512)
    max_length=512
)
print("Input IDs shape:", inputs["input_ids"].shape)
print("Attention Mask shape:", inputs["attention_mask"].shape)

# 3. Model Forward Pass (Disable autograd gradient computation)
with torch.no_grad():
    outputs = model(**inputs)

# 4. Extract unnormalized Logits
logits = outputs.logits  # shape: [batch_size=1, num_classes=2]
print("Raw Logits:", logits)

# 5. Postprocess: Softmax normalization over class dimension
probabilities = torch.nn.functional.softmax(logits, dim=-1)[0]
predicted_class_id = torch.argmax(probabilities).item()

# 6. Map ID to human-readable label using model configuration
label = model.config.id2label[predicted_class_id]
confidence = probabilities[predicted_class_id].item()

print(f"Prediction: {label} (Confidence: {confidence:.4f})")

from datasets import load_dataset

# 1. Load specific split directly
train_ds = load_dataset("imdb", split="train")
print("Number of training samples:", len(train_ds))
print("Feature schema:", train_ds.features)

# 2. Slicing syntax: first 1,000 examples only
small_test = load_dataset("imdb", split="test[:1000]")

# 3. Terabyte-scale streaming: stream without downloading full archive
stream_ds = load_dataset("imdb", split="train", streaming=True)
first_sample = next(iter(stream_ds))
print("Streamed sample label:", first_sample["label"])

---
title: Production Sentiment Classifier
emoji: 🚀
colorFrom: yellow
colorTo: amber
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
---

import gradio as gr
from transformers import pipeline

classifier = pipeline("sentiment-analysis", model="distilbert/distilbert-base-uncased-finetuned-sst-2-english")

def analyze_sentiment(text):
    result = classifier(text)[0]
    return {result["label"]: result["score"]}

demo = gr.Interface(
    fn=analyze_sentiment,
    inputs=gr.Textbox(lines=2, placeholder="Enter text to analyze..."),
    outputs=gr.Label(num_top_classes=2),
    title="Real-Time Sentiment Classifier"
)

if __name__ == "__main__":
    demo.launch()

# Good: Pinned to exact Git commit SHA
model = AutoModelForSequenceClassification.from_pretrained(
    "distilbert/distilbert-base-uncased-finetuned-sst-2-english",
    revision="714eb0fa89d4f8003548f7b7ca4b4d9e03f81ccb"
)
tokenizer = AutoTokenizer.from_pretrained(
    "distilbert/distilbert-base-uncased-finetuned-sst-2-english",
    revision="714eb0fa89d4f8003548f7b7ca4b4d9e03f81ccb"
)