import gradio as gr from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline from diffusers import StableDiffusionPipeline import torch import requests from PIL import Image import io # Define a dictionary of conversational models conversational_models = { "Qwen": "Qwen/QwQ-32B", "DeepSeek R1": "deepseek-ai/DeepSeek-R1", "Perplexity (R1 Post-trained)": "perplexity-ai/r1-1776", "Llama-Instruct by Meta": "meta-llama/Llama-3.2-3B-Instruct", "Mistral": "mistralai/Mistral-7B-v0.1", "Gemma": "google/gemma-2-2b-it", } # Define a dictionary of Text-to-Image models text_to_image_models = { "Stable Diffusion 3.5 Large": "stabilityai/stable-diffusion-3.5-large", "Stable Diffusion 1.4": "CompVis/stable-diffusion-v1-4", "Flux Dev": "black-forest-labs/FLUX.1-dev", } # Define a dictionary of Text-to-Speech models text_to_speech_models = { "Spark TTS": "SparkAudio/Spark-TTS-0.5B", } # Initialize tokenizers and models for conversational AI conversational_tokenizers = {} conversational_models_loaded = {} # Initialize pipelines for Text-to-Image text_to_image_pipelines = {} # Initialize pipelines for Text-to-Speech text_to_speech_pipelines = {} # Initialize pipelines for other tasks visual_qa_pipeline = pipeline("visual-question-answering", model="dandelin/vilt-b32-finetuned-vqa") document_qa_pipeline = pipeline("question-answering", model="deepset/roberta-base-squad2") image_classification_pipeline = pipeline("image-classification", model="facebook/detr-resnet-50") # This will be replaced object_detection_pipeline = pipeline("object-detection", model="facebook/detr-resnet-50") video_classification_pipeline = pipeline("video-classification", model="facebook/timesformer-base-finetuned-k400") # Removed text_to_3d_pipeline as it was causing issues keypoint_detection_pipeline = pipeline("keypoint-detection", model="microsoft/resnet50") # This will be replaced translation_pipeline = pipeline("translation_en_to_fr", model="Helsinki-NLP/opus-mt-en-fr") summarization_pipeline = pipeline("summarization", model="facebook/bart-large-cnn") text_to_audio_pipeline = pipeline("text-to-speech", model="julien-c/ljspeech_tts_train_tacotron2_raw_phn_tacotron_g2p_en_no_space") audio_classification_pipeline = pipeline("audio-classification", model="facebook/wav2vec2-base") def load_conversational_model(model_name): if model_name not in conversational_models_loaded: tokenizer = AutoTokenizer.from_pretrained(conversational_models[model_name]) model = AutoModelForCausalLM.from_pretrained(conversational_models[model_name]) conversational_tokenizers[model_name] = tokenizer conversational_models_loaded[model_name] = model return conversational_tokenizers[model_name], conversational_models_loaded[model_name] def chat(model_name, user_input, history=[]): tokenizer, model = load_conversational_model(model_name) # Encode the input input_ids = tokenizer.encode(user_input + tokenizer.eos_token, return_tensors="pt") # Generate a response with torch.no_grad(): output = model.generate(input_ids, max_length=150, pad_token_id=tokenizer.eos_token_id) response = tokenizer.decode(output[0], skip_special_tokens=True) # Clean up the response to remove the user input part response = response[len(user_input):].strip() # Append to chat history history.append((user_input, response)) return history, history def generate_image(model_name, prompt): if model_name not in text_to_image_pipelines: text_to_image_pipelines[model_name] = StableDiffusionPipeline.from_pretrained(text_to_image_models[model_name]) pipeline = text_to_image_pipelines[model_name] image = pipeline(prompt).images[0] return image def generate_speech(model_name, text): if model_name not in text_to_speech_pipelines: text_to_speech_pipelines[model_name] = pipeline("text-to-speech", model=text_to_speech_models[model_name]) pipeline = text_to_speech_pipelines[model_name] audio = pipeline(text) return audio["audio"] def visual_qa(image, question): result = visual_qa_pipeline(image, question) return result["answer"] def document_qa(document, question): result = document_qa_pipeline(question=question, context=document) return result["answer"] def image_classification(image): result = image_classification_pipeline(image) return result def object_detection(image): result = object_detection_pipeline(image) return result def video_classification(video): result = video_classification_pipeline(video) return result def keypoint_detection(image): result = keypoint_detection_pipeline(image) return result def translate_text(text): result = translation_pipeline(text) return result[0]["translation_text"] def summarize_text(text): result = summarization_pipeline(text) return result[0]["summary_text"] def text_to_audio(text): result = text_to_audio_pipeline(text) return result["audio"] def audio_classification(audio): result = audio_classification_pipeline(audio) return result # Define the Gradio interface with gr.Blocks() as demo: gr.Markdown("## Versatile AI Chatbot and Text-to-X Tasks") with gr.Tab("Conversational AI"): conversational_model_choice = gr.Dropdown(list(conversational_models.keys()), label="Choose a Conversational Model") conversational_chatbot = gr.Chatbot(label="Chat") conversational_message = gr.Textbox(label="Message") conversational_submit = gr.Button("Submit") conversational_submit.click(chat, inputs=[conversational_model_choice, conversational_message, conversational_chatbot], outputs=[conversational_chatbot, conversational_chatbot]) conversational_message.submit(chat, inputs=[conversational_model_choice, conversational_message, conversational_chatbot], outputs=[conversational_chatbot, conversational_chatbot]) with gr.Tab("Text-to-Image"): text_to_image_model_choice = gr.Dropdown(list(text_to_image_models.keys()), label="Choose a Text-to-Image Model") text_to_image_prompt = gr.Textbox(label="Prompt") text_to_image_generate = gr.Button("Generate Image") text_to_image_output = gr.Image(label="Generated Image") text_to_image_generate.click(generate_image, inputs=[text_to_image_model_choice, text_to_image_prompt], outputs=text_to_image_output) with gr.Tab("Text-to-Speech"): text_to_speech_model_choice = gr.Dropdown(list(text_to_speech_models.keys()), label="Choose a Text-to-Speech Model") text_to_speech_text = gr.Textbox(label="Text") text_to_speech_generate = gr.Button("Generate Speech") text_to_speech_output = gr.Audio(label="Generated Speech") text_to_speech_generate.click(generate_speech, inputs=[text_to_speech_model_choice, text_to_speech_text], outputs=text_to_speech_output) with gr.Tab("Visual Question Answering"): visual_qa_image = gr.Image(label="Upload Image") visual_qa_question = gr.Textbox(label="Question") visual_qa_generate = gr.Button("Answer") visual_qa_output = gr.Textbox(label="Answer") visual_qa_generate.click(visual_qa, inputs=[visual_qa_image, visual_qa_question], outputs=visual_qa_output) with gr.Tab("Document Question Answering"): document_qa_document = gr.Textbox(label="Document Text") document_qa_question = gr.Textbox(label="Question") document_qa_generate = gr.Button("Answer") document_qa_output = gr.Textbox(label="Answer") document_qa_generate.click(document_qa, inputs=[document_qa_document, document_qa_question], outputs=document_qa_output) with gr.Tab("Image Classification"): image_classification_image = gr.Image(label="Upload Image") image_classification_generate = gr.Button("Classify") image_classification_output = gr.Textbox(label="Classification Result") image_classification_generate.click(image_classification, inputs=image_classification_image, outputs=image_classification_output) with gr.Tab("Object Detection"): object_detection_image = gr.Image(label="Upload Image") object_detection_generate = gr.Button("Detect") object_detection_output = gr.Image(label="Detection Result") object_detection_generate.click(object_detection, inputs=object_detection_image, outputs=object_detection_output) with gr.Tab("Video Classification"): video_classification_video = gr.Video(label="Upload Video") video_classification_generate = gr.Button("Classify") video_classification_output = gr.Textbox(label="Classification Result") video_classification_generate.click(video_classification, inputs=video_classification_video, outputs=video_classification_output) # Removed Text-to-3D tab as it was causing issues with gr.Tab("Keypoint Detection"): keypoint_detection_image = gr.Image(label="Upload Image") keypoint_detection_generate = gr.Button("Detect Keypoints") keypoint_detection_output = gr.Image(label="Keypoint Detection Result") keypoint_detection_generate.click(keypoint_detection, inputs=keypoint_detection_image, outputs=keypoint_detection_output) with gr.Tab("Translation"): translate_text_text = gr.Textbox(label="Text") translate_text_generate = gr.Button("Translate") translate_text_output = gr.Textbox(label="Translated Text") translate_text_generate.click(translate_text, inputs=translate_text_text, outputs=translate_text_output) with gr.Tab("Summarization"): summarize_text_text = gr.Textbox(label="Text") summarize_text_generate = gr.Button("Summarize") summarize_text_output = gr.Textbox(label="Summary") summarize_text_generate.click(summarize_text, inputs=summarize_text_text, outputs=summarize_text_output) with gr.Tab("Text-to-Audio"): text_to_audio_text = gr.Textbox(label="Text") text_to_audio_generate = gr.Button("Generate Audio") text_to_audio_output = gr.Audio(label="Generated Audio") text_to_audio_generate.click(text_to_audio, inputs=text_to_audio_text, outputs=text_to_audio_output) with gr.Tab("Audio Classification"): audio_classification_audio = gr.Audio(label="Upload Audio") audio_classification_generate = gr.Button("Classify") audio_classification_output = gr.Textbox(label="Classification Result") audio_classification_generate.click(audio_classification, inputs=audio_classification_audio, outputs=audio_classification_output) # Launch the Gradio interface demo.launch()