app.py

from fastapi import FastAPI, UploadFile, File
from transformers import pipeline
from fastai.vision.all import *
from PIL import Image

# NOTE - we configure docs_url to serve the interactive Docs at the root path
# of the app. This way, we can use the docs as a landing page for the app on Spaces.
app = FastAPI(docs_url="/")

pipe = pipeline("text2text-generation", model="google/flan-t5-small")
categories = ('Heart', 'Oblong', 'Oval', 'Round', 'Square')
learn = load_learner('model.pkl')

@app.get("/generate")
def generate(text: str):
    """
    Using the text2text-generation pipeline from `transformers`, generate text
    from the given input text. The model used is `google/flan-t5-small`, which
    can be found [here](https://huggingface.co/google/flan-t5-small).
    """
    output = pipe(text)
    return {"output": output[0]["generated_text"]}

@app.post("/face-analyse")
async def face_analyse(file: UploadFile = File(...)):
    # Read the uploaded file content
    request_object_content = await file.read()
    
    try:
        # Attempt to open the image
        img = Image.open(io.BytesIO(request_object_content))
    except Exception as e:
        return {"error": "Failed to open the image file. Make sure it is a valid image file."}

    # Check if img is None or not
    if img is None:
        return {"error": "Failed to open the image file."}
    
    try:
        # Resize the image to 300x300 pixels
        img = img.resize((300, 300))
    except Exception as e:
        return {"error": "Failed to resize the image."}

    try:
        # Assuming 'learn' is your image classifier model
        pred, idx, probs = learn.predict(img)
    except Exception as e:
        return {"error": "Failed to make predictions."}

    # Assuming categories is a list of category labels
    return dict(zip(categories, map(float, probs)))

    # Initialize the Meta-Llama-3-70B-Instruct pipeline
llama_model_id = "meta-llama/Meta-Llama-3-70B-Instruct"
llama_pipeline = pipeline(
    "text-generation",
    model=llama_model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

@app.post("/frame-details")
def frame_details(text: str):
    """
    Extract structured information from a given text about frames using the
    Meta-Llama-3-70B-Instruct model. The model will output details like price, color, etc.
    """
    messages = [
        {"role": "system", "content": "You are an api chatbot for frames and glasses who always responds with only a json. Extract the infomation given into a structured json for frame details"},
        {"role": "user", "content": text},
    ]

    terminators = [
        llama_pipeline.tokenizer.eos_token_id,
        llama_pipeline.tokenizer.convert_tokens_to_ids("")
    ]

    outputs = llama_pipeline(
        messages,
        max_new_tokens=256,
        eos_token_id=terminators,
        do_sample=True,
        temperature=0.6,
        top_p=0.9,
    )
    
    # Extract the last generated text from the output
    generated_text = outputs[0]["generated_text"]
    
    # Parse the generated text to extract structured information (this is an example and should be customized)
    # Here, you would add your own logic to parse the generated text
    # For now, we'll assume the generated text is in JSON format
    try:
        extracted_info = eval(generated_text)  # It's recommended to use `json.loads` in a real application
    except Exception as e:
        return {"error": "Failed to parse the generated text."}

    return extracted_info

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=8000)