app.py

import gradio as gr
import torch
import torch.nn.functional as F
import torch.nn as nn
from transformers import AutoProcessor, AutoModel
from peft import PeftModel
from PIL import Image

class ClassificationHead(nn.Module):
    def __init__(self, input_dim):
        super().__init__()
        self.linear = nn.Linear(input_dim, 2)

    def forward(self, x):
        return self.linear(x)

def load_model():
    device = torch.device("cpu")

    base_model = AutoModel.from_pretrained(
        "google/siglip-so400m-patch14-384",
        device_map="cpu",
        torch_dtype=torch.float32,
        attn_implementation="sdpa"
    ).vision_model

    model = PeftModel.from_pretrained(base_model, "fumo_lora", local_files_only=True)
    processor = AutoProcessor.from_pretrained("google/siglip-so400m-patch14-384")

    head = ClassificationHead(1152)
    head.load_state_dict(torch.load("fumo_lora/classification_head.pth", weights_only=True, map_location="cpu"))

    model.eval()
    head.eval()

    return model, processor, head, device

model, processor, head, device = load_model()

def predict_image(image):
    if image is None:
        return "Please provide an image."

    try:
        # Process image
        inputs = processor(images=image, return_tensors="pt")

        with torch.no_grad():
            outputs = model(
                pixel_values=inputs.pixel_values.to(device, dtype=torch.float32),
            )
            pooled = outputs.last_hidden_state.mean(dim=1)
            logits = head(pooled)
            prob = F.softmax(logits, dim=1)

        fumo_prob = prob[0, 1].item()
        not_fumo_prob = prob[0, 0].item()

        result = f"Results:\n"
        result += f"Fumo probability: {fumo_prob:.3f}\n"
        result += f"Not fumo probability: {not_fumo_prob:.3f}\n"
        result += f"\nVerdict: {'FUMO!' if fumo_prob > 0.5 else 'Not a fumo'}"

        return result

    except Exception as e:
        return f"Error: {str(e)}"

htmlhead = """
<script>
function onLoad() {
    setTimeout(() => {
        const buttons = [...document.querySelectorAll("button")].filter(v => v.innerText.includes("Flag as"));
        buttons.forEach(v => v.disabled = true);

        const submit = [...document.querySelectorAll("button")].filter(v => v.innerText.includes("Submit"))[0];

        submit.addEventListener("click", function() {
            buttons.forEach(v => v.disabled = false);
        });
    }, 1500);
}

if (document.readyState === 'complete') {
    onLoad();
} else {
    window.addEventListener('load', onLoad);
}
</script>
"""

# Create Gradio interface
demo = gr.Interface(
    fn=predict_image,
    inputs=gr.Image(type="pil", width=384, height=384),
    outputs=gr.Textbox(),
    title="Fumo Classifier (LoRA)",
    description="Drop an image to check if it's a Fumo!",
    examples=["examples/fumo1.jpg", "examples/fumo2.jpg", "examples/no_fumo1.jpg", "examples/no_fumo2.jpg", "examples/no_fumo3.png"],
    flagging_mode="manual",
    flagging_options=["Correct 👍", "Incorrect 👎"],
    head=htmlhead,
)

if __name__== "__main__":
    has_bf16 = torch.cuda.is_bf16_supported() if torch.cuda.is_available() else False
    # or for CPU:
    has_bf16_cpu = torch.cpu.is_bf16_supported() if hasattr(torch.cpu, 'is_bf16_supported') else False
    print(f"BF16 support: {has_bf16} (GPU), {has_bf16_cpu} (CPU)")
    
    demo.launch()