app.py

# Set random seeds for reproducibility
import random
import numpy as np
import torch
from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    DataCollatorWithPadding,
    TrainingArguments,
    Trainer,
    EarlyStoppingCallback
)
from transformers import TextClassificationPipeline
from sklearn.metrics import accuracy_score, f1_score
from transformers_interpret import SequenceClassificationExplainer
from transformers import pipeline
import gradio as gr

SEED = 42
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
if torch.cuda.is_available():
    torch.cuda.manual_seed_all(SEED)

USE_MPS = torch.backends.mps.is_available()
device = torch.device("mps" if USE_MPS else "cpu")
print("Using device:", device)

# Load the ag_news dataset
raw = load_dataset("SetFit/ag_news")
print(raw)


# Load BERT tokenizer
MODEL_NAME = "bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

# Tokenization function
def tokenize_fn(examples):
    return tokenizer(examples["text"], truncation=True, max_length=128)

cols_to_remove = [c for c in raw["train"].column_names if c not in ("label",)]

# Apply tokenization to the dataset
tokenized = raw.map(tokenize_fn, batched=True, remove_columns=cols_to_remove)


# Remove original text column to avoid issues during batching
if "text" in tokenized["train"].column_names:
    tokenized = tokenized.remove_columns(["text"])

# Set dataset format to PyTorch tensors
tokenized.set_format("torch")

# Shuffle and split the training dataset to create a validation set
train_dataset = tokenized["train"].shuffle(seed=SEED)
val_split = train_dataset.train_test_split(test_size=5000, seed=SEED)
train_dataset = val_split["train"]
eval_dataset = val_split["test"]

print(train_dataset)

# Load pre-trained BERT model for sequence classification
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME, num_labels=4)

# Create a data collator that dynamically pads input sequences in each batch
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)


# Define a metrics computation function using scikit-learn
def compute_metrics(eval_pred):
    logits, labels = eval_pred
    # Convert logits to predicted class indices
    preds = np.argmax(logits, axis=-1)

    # Compute accuracy and F1 score using scikit-learn
    acc = accuracy_score(labels, preds)
    f1 = f1_score(labels, preds, average='macro')

    return {"accuracy": acc, "f1_macro": f1}

# Define training arguments
training_args = TrainingArguments(
    output_dir="./results",
    eval_strategy="epoch",
    save_strategy="epoch",
    logging_strategy="epoch",
    #report_to=[],  # <- disable all integrations (no wandb, no tensorboard)
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    num_train_epochs=3,
    learning_rate=2e-5,
    weight_decay=0.1,
    warmup_steps=100,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",
    greater_is_better=False,
    save_total_limit=3,
    fp16=torch.cuda.is_available(),
    dataloader_drop_last=False,
    gradient_accumulation_steps=1,
    seed=SEED,
)

# Create Trainer instance with early stopping
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    tokenizer=tokenizer,
    data_collator=data_collator,
    compute_metrics=compute_metrics,
    callbacks=[EarlyStoppingCallback(early_stopping_patience=2)],
)

# Start model training
trainer.train()

# Save the fine-tuned model
trainer.save_model('my-fine-tuned-bert')

# Save the tokenizer
tokenizer.save_pretrained('my-fine-tuned-bert')

# Load the fine-tuned model and tokenizer
new_model = AutoModelForSequenceClassification.from_pretrained('my-fine-tuned-bert')
new_tokenizer = AutoTokenizer.from_pretrained('my-fine-tuned-bert')

# Create a text classification pipeline
classifier = TextClassificationPipeline(
    model=new_model, 
    tokenizer=new_tokenizer,     )

# Define label mapping
label_mapping = {
    0: 'World',
    1: 'Sports',
    2: 'Business',
    3: 'Sci/Tech'
}

# Test the classifier on a sample sentence
sample_text = "This movie was good"
result = classifier(sample_text)

# Map the predicted label to a meaningful sentiment
mapped_result = {
    'label': label_mapping[int(result[0]['label'].split('_')[1])],
    'score': result[0]['score']
}

print(mapped_result)

MODEL_ID = "my-fine-tuned-bert"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
explainer = SequenceClassificationExplainer(model=model, tokenizer=tokenizer)

label_names = {0: "World", 1: "Sports", 2: "Business", 3: "Sci/Tech"}

device = 0 if torch.cuda.is_available() else -1
clf = pipeline("text-classification", model=model, tokenizer=tokenizer, top_k=None, device=device)

def predict(text: str):
    text = (text or "").strip()
    if not text:
        return {}
    out = clf(text, truncation=True)
    if isinstance(out, list) and isinstance(out[0], list):
        out = out[0]
    results = {}
    for o in sorted(out, key=lambda x: -x["score"]):
        idx = int(o["label"].split("_")[1])
        results[label_names[idx]] = float(o["score"])
    return results

# Build script-free HTML so it renders in Gradio pages
def explain_html(text: str) -> str:
    text = (text or "").strip()
    if not text:
        return "<i>Enter text to see highlighted words.</i>"
    atts = explainer(text)
    toks = [t for t, _ in atts]
    scores = np.abs([s for _, s in atts])
    smin, smax = float(np.min(scores)), float(np.max(scores))
    scores = (scores - smin) / (smax - smin + 1e-8)
    spans = [
        f"<span style='background: rgba(255,0,0,{0.15+0.85*s:.2f});"
        f"padding:2px 3px; margin:1px; border-radius:4px; display:inline-block'>{tok}</span>"
        for tok, s in zip(toks, scores)
    ]
    return "<div style='line-height:2'>" + " ".join(spans) + "</div>"

def predict_and_explain(text: str):
    return predict(text), explain_html(text)

demo = gr.Interface(
    fn=predict_and_explain,
    inputs=gr.Textbox(lines=3, label="Enter news headline"),
    outputs=[
        gr.Label(num_top_classes=4, label="Predicted topic"),
        gr.HTML(label="Important-word highlights"),
    ],
    title="AG News Topic Classifier (BERT-base)",
    description="Shows predicted topic and highlights words that influenced the decision."
)

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