app.py

from PIL import Image
import gradio as gr

import torch
import torch.nn as nn
from transformers import AutoTokenizer, pipeline
from transformers import AutoModelForCausalLM
from torchvision import transforms
from transformers import CLIPProcessor, CLIPModel

from model import build_mlp_vector_projector


device = "cpu"
# Load the CLIP model and processor
clip_model_name = "openai/clip-vit-base-patch16"
clip_model = CLIPModel.from_pretrained(clip_model_name).to(device)
clip_processor = CLIPProcessor.from_pretrained(clip_model_name)

clip_transform = transforms.Compose(
    [
        transforms.Resize((224, 224)),
        transforms.ToTensor()
    ]
)


def process_image(img_path):
    image = Image.open(img_path).convert("RGB")
    image = clip_transform(image)
    inputs = clip_processor(text=[""], images=image,
                            return_tensors="pt", padding=True)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    img_embedding = clip_model(**inputs).image_embeds
    img_proj_head = build_mlp_vector_projector().to(device)
    img_proj_head.load_state_dict(torch.load(
        'stage_2_proj_head_v3.pth', map_location=torch.device(device)))
    img_tokens = img_proj_head(img_embedding)
    return img_tokens


phi_model_name = "microsoft/phi-2"
text_tokenizer = AutoTokenizer.from_pretrained(
    phi_model_name, trust_remote_code=True)
with torch.no_grad():
    base_phi2_text = AutoModelForCausalLM.from_pretrained(
        phi_model_name, trust_remote_code=True,
        device_map="auto", torch_dtype=torch.float16
    )
    tuned_phi2 = AutoModelForCausalLM.from_pretrained(
        "stage2_adaptor", trust_remote_code=True,
    ).to("cpu")
print("phi2 model loaded")

audio_model_name = "openai/whisper-small"
audio_pipe = pipeline(
    task="automatic-speech-recognition",
    model=audio_model_name,
    chunk_length_s=30,
    device=device)


def process_text(text, count):
    inputs = text_tokenizer.encode(text, return_tensors="pt")
    input_embeds = tuned_phi2.get_submodule(
            'model.embed_tokens')(inputs).to(device)
    prediction = text_tokenizer.batch_decode(
         tuned_phi2.generate(
                inputs_embeds=input_embeds,
                max_new_tokens=30,
                bos_token_id=text_tokenizer.bos_token_id,
                eos_token_id=text_tokenizer.eos_token_id,
                pad_token_id=text_tokenizer.pad_token_id
        )
    )
    return prediction[0].rstrip('<|endoftext|>').rstrip("\n")


def process_audio(audio):
    if audio is None:
        raise gr.Error(
            "Please provide an audio file or record your input"
        )

    text = audio_pipe(
        audio,
        batch_size=8,
        generate_kwargs={"task": "transcribe"},
        return_timestamps=True
    )["text"]
    return text


def generate_response(image, audio, text, count):
    count = int(count)
    overall_input = ""
    if audio:
        overall_input = process_audio(audio)
    if text:
        overall_input = text + overall_input
    if image:
        img_tokens = process_image(image)
        overall_input = "Question: " + overall_input + "Answer:" 
        q_tokens = text_tokenizer.encode(
            overall_input,
            return_tensors='pt').to(device)
        question_token_embeddings = tuned_phi2.get_submodule(
            'model.embed_tokens')(q_tokens).to(device)
        inputs = torch.concat(
            (img_tokens.unsqueeze(0), question_token_embeddings),
            axis=-2).to(device)
        prediction = text_tokenizer.batch_decode(
            tuned_phi2.generate(
                inputs_embeds=inputs,
                max_new_tokens=30,
                bos_token_id=text_tokenizer.bos_token_id,
                eos_token_id=text_tokenizer.eos_token_id,
                pad_token_id=text_tokenizer.pad_token_id
            )
        )
        return prediction[0].rstrip('<|endoftext|>').rstrip("\n")
    else:
        return process_text(overall_input, count)


with gr.Blocks() as demo:
    gr.Markdown("# **AnyModeAssistant**")
    gr.Markdown("Use any mode text/image/audio to interact with AI assistant")
    with gr.Row():
        with gr.Column(scale=4):
            with gr.Row("Text"):
                text_input = gr.Textbox(placeholder="Enter your question here",
                                        label="Input")
            with gr.Row():
                image_input = gr.Image(type="filepath")

            with gr.Row("Audio mode"):
                audio_input = gr.Audio(type="filepath")

            with gr.Row("Image"):
                response_count = gr.Textbox(
                    placeholder="Number of tokens to respond",
                    value=20,
                    label="Count")
        with gr.Column(scale=2):
            response = gr.Textbox(label="AI Response")
    with gr.Row():
        submit_button = gr.Button("Submit")
        submit_button.click(generate_response,
                            inputs=[
                                image_input, audio_input,
                                text_input, response_count
                            ],
                            outputs=response)

    gr.Examples(
        examples=[
            ["dog_man_forest.jpg", "audio.m4a", "Is there a dog present in the image?"],
        ],
        inputs=[input_image, audio_input, text_input, response_count],
        outputs=[response],
        fn=generate_response,
    )

demo.launch(share=True)