app.py

import subprocess
# Installing flash_attn
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)

import gradio as gr
from PIL import Image
from transformers import AutoModelForCausalLM
from transformers import AutoProcessor
from transformers import TextIteratorStreamer
import time
from threading import Thread
import torch
import spaces

model_id = "microsoft/Phi-3-vision-128k-instruct"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="cuda", trust_remote_code=True, torch_dtype="auto")
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model.to("cuda:0")

# Enhanced Placeholder HTML with instructions and centralization
PLACEHOLDER = """
<div style="padding: 30px; text-align: center; display: flex; flex-direction: column; align-items: center; justify-content: center; background-image: url('https://huggingface.co/spaces/simonraj/PersonalTrainer-Arnold/blob/main/fitness_coach_app_resized.jpg'); background-size: cover; background-position: center; width: 100%; height: 100vh;">
   <div style="background-color: rgba(255, 255, 255, 0.8); padding: 20px; border-radius: 10px; width: 80%; max-width: 550px; text-align: center;">
       <h1 style="font-size: 32px; margin-bottom: 10px; color: black;">Get Ripped with Arnold's AI Coach</h1>
       <p style="font-size: 20px; margin-bottom: 10px; color: black;">Welcome to the ultimate fitness companion! 💪</p>
       <ul style="text-align: left; font-size: 18px; list-style: none; padding: 0; color: black;">
          <li>📸 <strong>Upload</strong> a photo of your exercise.</li>
          <li>⚡ <strong>Get instant feedback</strong> to perfect your form.</li>
          <li>🔥 <strong>Improve your workouts</strong> with expert tips!</li>
       </ul>
   </div>
</div>
"""

@spaces.GPU
def bot_streaming(message, history):
    print(f'message is - {message}')
    print(f'history is - {history}')
    if message["files"]:
        if type(message["files"][-1]) == dict:
            image = message["files"][-1]["path"]
        else:
            image = message["files"][-1]
    else:
        for hist in history:
            if type(hist[0]) == tuple:
                image = hist[0][0]
    try:
        if image is None:
            raise gr.Error("You need to upload an image for Phi3-Vision to work. Close the error and try again with an Image.")
    except NameError:
        raise gr.Error("You need to upload an image for Phi3-Vision to work. Close the error and try again with an Image.")

    conversation = []
    flag = False
    for user, assistant in history:
        if assistant is None:
            flag = True
            conversation.extend([{"role": "user", "content": ""}])
            continue
        if flag == True:
            conversation[0]['content'] = f"\n{user}"
            conversation.extend([{"role": "assistant", "content": assistant}])
            flag = False
            continue
        conversation.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])

    if len(history) == 0:
        conversation.append({"role": "user", "content": f"\n{message['text']}"})
    else:
        conversation.append({"role": "user", "content": message['text']})
    print(f"prompt is -\n{conversation}")
    prompt = processor.tokenizer.apply_chat_template(conversation, tokenize=False, add_generation_prompt=True)
    image = Image.open(image)
    inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")

    # Custom prompt to ensure responses are in Arnold's style
    system_prompt = (
        "As Arnold Schwarzenegger, analyze the image to identify the exercise being performed. "
        "Provide detailed coaching tips to improve the form, focusing on posture and common errors. "
        "Use motivational and energetic language. If the image does not show an exercise, respond with: "
        "'What are you doing? This is no time for games! Upload a real exercise picture and let's pump it up!'"
    )

    streamer = TextIteratorStreamer(processor, **{"skip_special_tokens": True, "skip_prompt": True, 'clean_up_tokenization_spaces': False,})
    generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=280, do_sample=False, temperature=0.0, eos_token_id=processor.tokenizer.eos_token_id,)

    thread = Thread(target=model.generate, kwargs=generation_kwargs, args=(system_prompt,))
    thread.start()

    buffer = ""
    for new_text in streamer:
        buffer += new_text
        yield buffer


chatbot = gr.Chatbot(scale=1, placeholder=PLACEHOLDER)
chat_input = gr.MultimodalTextbox(interactive=True, file_types=["image"], placeholder="Enter message or upload file...", show_label=False)
with gr.Blocks(fill_height=True,) as demo:
    gr.ChatInterface(
        fn=bot_streaming,
        title="Get Ripped with Arnold's AI Coach",
        examples=[
            {"text": "Identify and provide coaching cues for this exercise.", "files": ["./squat.jpg"]},
            {"text": "What improvements can I make?", "files": ["./pushup.jpg"]},
            {"text": "How is my form?", "files": ["./plank.jpg"]},
            {"text": "Give me some tips to improve my deadlift.", "files": ["./deadlift.jpg"]}
        ],
        description="Welcome to the ultimate fitness companion! 💪\nUpload a photo of your exercise and get instant feedback to perfect your form. Improve your workouts with expert tips!",
        stop_btn="Stop Generation",
        multimodal=True,
        textbox=chat_input,
        chatbot=chatbot,
        cache_examples=False,
        examples_per_page=3
    )

demo.queue()
demo.launch(debug=True, quiet=True)