app.py

import warnings
warnings.simplefilter('ignore')
import numpy as np
import torch
import torch.nn as nn
from transformers import DistilBertTokenizer, DistilBertModel
import logging
logging.basicConfig(level=logging.ERROR)
from torch import cuda
import gradio as gr

def classify(sentence):
	output = ""
	class DistilBERTClass(nn.Module):
		def __init__(self, num_intents):
			super(DistilBERTClass, self).__init__()
			self.l1 = DistilBertModel.from_pretrained("distilbert-base-uncased")
			self.fc1 = nn.Sequential(
				nn.Linear(768, 64),
				nn.BatchNorm1d(64),
				nn.ReLU(),
			)
			self.fc2 = nn.Sequential(
				nn.Linear(64, num_intents)
			)

		def forward(self, input_ids, attention_mask):
			output_1 = self.l1(input_ids=input_ids, attention_mask=attention_mask)
			hidden_state = output_1[0]
			pooler = hidden_state[:, 0]
			pooler = self.fc1(pooler)
			output = self.fc2(pooler)
			return output

	user_intents = ['initial_query', 'greeting', 'add_filter', 'remove_filter', 'continue', 'accept_response', 'reject_response']
	musical_attributes = ['track', 'artist', 'year', 'popularity', 'culture', 'similar_track', 'similar_artist', 'user', 'theme', 'mood', 'genre', 'instrument', 'vocal', 'tempo']
	intents_dict = {"user": user_intents, "music": musical_attributes}
	num_intents_dict = {'user': 7, 'music': 14}

	device = 'cuda:0' if cuda.is_available() else 'cpu'
	tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")

	for data_type in ["user", "music"]:
		num_intents = num_intents_dict[data_type]

		model = DistilBERTClass(num_intents)
		model.load_state_dict(torch.load(f"./models/{data_type}_finetune_model.pth", map_location=torch.device('cpu')))
		model.to(device)

		# Tokenize the input sentence
		inputs = tokenizer.encode_plus(
			sentence,
			None,
			add_special_tokens=True,
			max_length=128,
			pad_to_max_length=True,
			return_token_type_ids=False,
			return_attention_mask=True,
			truncation=True
		)

		input_ids = torch.tensor(inputs['input_ids']).unsqueeze(0).to(device)
		attention_mask = torch.tensor(inputs['attention_mask']).unsqueeze(0).to(device)

		model.eval()
		with torch.no_grad():
			outputs = model(input_ids, attention_mask)
			probability_outputs = torch.sigmoid(outputs).cpu().detach().numpy()

		binary_outputs = (probability_outputs >= 0.5)
		binary_outputs[np.all(binary_outputs == False, axis=1), -1] = True

		intents = intents_dict[data_type]
		predicted_intents = [intent for i, intent in enumerate(intents) if binary_outputs[0][i] == 1]

		if data_type=="user":
			output += f"User Intents: {predicted_intents}\n"
		else:
			output += f"Musical Attributes: {predicted_intents}\n"
	return output

title = "User Intents and Musical Attributes Classifier"
description = """
You can engage in a conversation with the music recommendation system, imagining a situation where it recommends music to you. The model will then predict the intents and musical attributes based on the sentence you provide.
<img src="https://github.com/user-attachments/assets/a8bfb1dc-856b-4f85-82dd-510cddcc2aeb" width=400px>
"""
article = "For more information, visit [Github Repository.](https://github.com/DaeyongKwon98/Intent-Classification/tree/main)"

demo = gr.Interface(
    fn=classify,
    inputs="text",
    outputs="text",
    title=title,
    description=description,
    article=article,
    examples=[["Hi, I need a playlist of rock songs to listen when I exercise."], ["I love Ariana Grande! Give me more."], ["I think these are too fast for me."]],
)

demo.launch()