import streamlit as st from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer from diffusers import StableDiffusionPipeline import torch from PIL import Image import librosa import tempfile import os # Configuração da página st.set_page_config(page_title="Demo Multi-Modal AI", page_icon="🤖", layout="wide") # -------- Cache de modelos -------- @st.cache_resource(show_spinner=False) def load_model(model_key): device = torch.device("cuda" if torch.cuda.is_available() else "cpu") cache_dir = "model_cache" os.makedirs(cache_dir, exist_ok=True) if model_key == 'sentiment_analysis': return pipeline("sentiment-analysis", model="cardiffnlp/twitter-roberta-base-sentiment-latest", device=device, cache_dir=cache_dir) elif model_key == 'text_classification': return pipeline("text-classification", model="distilbert-base-uncased-finetuned-sst-2-english", device=device, cache_dir=cache_dir) elif model_key == 'summarization': return pipeline("summarization", model="facebook/bart-large-cnn", device=device, max_length=150, min_length=30, cache_dir=cache_dir) elif model_key == 'question_answering': return pipeline("question-answering", model="deepset/roberta-base-squad2", device=device, cache_dir=cache_dir) elif model_key == 'translation': return pipeline("translation", model="Helsinki-NLP/opus-mt-tc-big-en-pt", device=device, cache_dir=cache_dir) elif model_key == 'text_generation': tokenizer = AutoTokenizer.from_pretrained("gpt2", cache_dir=cache_dir) model = AutoModelForCausalLM.from_pretrained("gpt2", cache_dir=cache_dir) model.config.pad_token_id = model.config.eos_token_id return pipeline("text-generation", model=model, tokenizer=tokenizer, device=device) elif model_key == 'ner': return pipeline("ner", model="dbmdz/bert-large-cased-finetuned-conll03-english", device=device, aggregation_strategy="simple", cache_dir=cache_dir) elif model_key == 'image_classification': return pipeline("image-classification", model="google/vit-base-patch16-224", device=device, cache_dir=cache_dir) elif model_key == 'object_detection': return pipeline("object-detection", model="facebook/detr-resnet-50", device=device, cache_dir=cache_dir) elif model_key == 'speech_to_text': return pipeline("automatic-speech-recognition", model="openai/whisper-base", device=device, cache_dir=cache_dir) elif model_key == 'audio_classification': return pipeline("audio-classification", model="superb/hubert-base-superb-er", device=device, cache_dir=cache_dir) elif model_key == 'text_to_image': return StableDiffusionPipeline.from_pretrained( "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, use_safetensors=True, safety_checker=None, cache_dir=cache_dir ) # -------- Funções auxiliares -------- def process_audio_file(audio_file): with tempfile.NamedTemporaryFile(delete=False, suffix=os.path.splitext(audio_file.name)[1]) as tmp_file: tmp_file.write(audio_file.read()) tmp_file_path = tmp_file.name audio_array, sr = librosa.load(tmp_file_path, sr=16000) os.unlink(tmp_file_path) return audio_array def process_image_file(image_file): image = Image.open(image_file) if image.mode != 'RGB': image = image.convert('RGB') return image def display_results(result, model_key, input_text=None): if model_key == 'summarization': st.subheader("📝 Resumo") if input_text: st.markdown("**Texto Original:**") st.write(input_text) st.info(result[0]['summary_text']) elif model_key == 'translation': st.subheader("🌍 Tradução") st.success(result[0]['translation_text']) elif model_key in ['sentiment_analysis', 'text_classification']: st.subheader("📊 Resultados") for res in result: st.write(f"- **{res['label']}**: {res['score']:.2%}") elif model_key == 'ner': st.subheader("🔍 Entidades Reconhecidas") for entity in result: st.write(f"- **{entity['word']}**: {entity['entity_group']} ({entity['score']:.2%})") elif model_key == 'text_generation': st.subheader("🧠 Texto Gerado") st.write(result[0]['generated_text']) elif model_key == 'image_classification': st.subheader("🏷️ Classificação de Imagem") for res in result[:5]: st.write(f"- **{res['label']}**: {res['score']:.2%}") elif model_key == 'object_detection': st.subheader("📦 Objetos Detectados") for obj in result: st.write(f"- {obj['label']} ({obj['score']:.2%})") elif model_key == 'speech_to_text': st.subheader("🔈 Transcrição de Áudio") st.success(result['text']) elif model_key == 'audio_classification': st.subheader("🎧 Classificação de Áudio") top_emotion = result[0] st.write(f"**Emoção detectada**: {top_emotion['label']} ({top_emotion['score']:.2%})") elif model_key == 'text_to_image': st.subheader("🎨 Imagem Gerada") st.image(result[0], caption="Imagem gerada a partir do texto") # -------- Casos de uso -------- use_cases = { 'sentiment_analysis': "A entrega foi super rápida, adorei!", 'text_classification': "Estou insatisfeito com o produto", 'summarization': "A empresa XYZ reportou um crescimento de 15% no último trimestre...", 'question_answering': { 'context': "O produto X tem garantia de 2 anos e pode ser configurado via app em 5 minutos.", 'question': "Qual é o tempo de garantia do produto X?" }, 'translation': "Our product ensures high performance", 'ner': "Microsoft assinou um contrato com a empresa XYZ em Nova York.", 'text_generation': "Era uma vez um robô que", 'speech_to_text': None, 'audio_classification': None, 'image_classification': None, 'object_detection': None, 'text_to_image': "Um carro futurista voando sobre Lisboa" } # -------- Interface -------- st.title("🤖 Demo Multi-Modal AI") model_key = st.selectbox("Escolha o modelo para testar:", list(use_cases.keys())) model = load_model(model_key) if model_key in ['sentiment_analysis', 'text_classification', 'summarization', 'translation', 'text_generation', 'ner']: input_text = st.text_area("Insira texto:", value=use_cases[model_key] if isinstance(use_cases[model_key], str) else "") if st.button("Executar"): if model_key == 'question_answering': result = model(question=use_cases['question_answering']['question'], context=use_cases['question_answering']['context']) else: result = model(input_text) display_results(result, model_key, input_text=input_text) elif model_key in ['speech_to_text', 'audio_classification']: audio_file = st.file_uploader("Carregue um arquivo de áudio", type=['wav','mp3','flac','m4a']) if audio_file and st.button("Executar"): audio_data = process_audio_file(audio_file) result = model(audio_file) display_results(result, model_key) elif model_key in ['image_classification', 'object_detection', 'text_to_image']: uploaded_file = st.file_uploader("Carregue uma imagem (ou deixe vazio para gerar)", type=['jpg','jpeg','png']) prompt = st.text_input("Prompt para gerar imagem (apenas text_to_image):", value=use_cases['text_to_image'] if model_key=='text_to_image' else "") if st.button("Executar"): if model_key == 'text_to_image': result = [model(prompt).images[0]] elif uploaded_file: image = process_image_file(uploaded_file) result = model(image) else: st.warning("Carregue uma imagem ou insira prompt para gerar.") result = None if result: display_results(result, model_key)