app.py

import gradio as gr
import torch
import torch.nn.functional as F
import random

from captum.attr import LayerConductance, LayerIntegratedGradients
from visualization import VisualizationDataRecord, visualize_text

from transformers import DistilBertTokenizerFast
from transformers import DistilBertForSequenceClassification
tokenizer = DistilBertTokenizerFast.from_pretrained('distilbert-base-uncased')
model = DistilBertForSequenceClassification.from_pretrained("./checkpoint-57000")

def summarize_attributions(attributions):
    attributions = attributions.sum(dim=-1).squeeze(0)
    attributions = attributions / torch.norm(attributions)
    return attributions

def forward_func(inputs, attention_mask):
    outputs = model(input_ids=inputs, attention_mask=attention_mask)
    logits = F.softmax(outputs.logits)
    return logits

def model_inference(text):
    encodings = tokenizer(text, truncation=True, padding=True, return_tensors="pt",)

    outputs = model(**encodings)
    
    logits = F.softmax(outputs.logits)
    logits = logits[0].detach().numpy()
    labels = {"human_written": float(logits[0]), "AI_generated": float(logits[1])}

    ref_input_ids = torch.zeros_like(encodings['input_ids'])
    lig = LayerIntegratedGradients(forward_func, model.distilbert.embeddings)
    attributions_start, delta_start = lig.attribute(inputs=encodings['input_ids'],
                                    baselines=ref_input_ids,
                                    additional_forward_args=(encodings['attention_mask']),
                                    return_convergence_delta=True,
                                    target=0)

    attributions_start_sum = summarize_attributions(attributions_start)
    start_scores = forward_func(encodings['input_ids'], encodings['attention_mask'])
    indices = encodings['input_ids'][0].detach().tolist()
    all_tokens = tokenizer.convert_ids_to_tokens(indices)
    ground_truth_start_ind = encodings['input_ids'][0][0].numpy()

    start_position_vis = VisualizationDataRecord(
                            attributions_start_sum,
                            torch.max(torch.softmax(start_scores[0], dim=0)),
                            torch.argmax(start_scores),
                            torch.argmax(start_scores),
                            str(ground_truth_start_ind),
                            attributions_start_sum.sum(),       
                            all_tokens,
                            delta_start)


    print('\033[1m', 'Visualizations For Start Position', '\033[0m')
    img = visualize_text([start_position_vis])
    
    html = (
        ""
        + img
        + ""
    )
    

    return labels, html

input_text = gr.Textbox(placeholder="Enter sentence here...")
output = [gr.outputs.Label(num_top_classes=2), "html"] 
#output = [gr.outputs.Label(num_top_classes=2), "highlight"] 
#output = [gr.outputs.Label(num_top_classes=2), "image"] 

# label = gr.outputs.Label(num_top_classes=2)

demo = gr.Interface(
    model_inference,
    input_text,
    output,
    examples=[
        ["Is this text written by a human?"],
        ["Or is this text generated by an AI model?"],
    ],
)

demo.launch()