# custom_node_furniture_mask.py by StyleSpace (and GPT4)
import torch
import torchvision.transforms as T
from torchvision.models.segmentation import deeplabv3_resnet50
from PIL import Image

class FurnitureMaskNode:
    def __init__(self):
        self.model = deeplabv3_resnet50(pretrained=True).eval()
        self.transforms = T.Compose([
            T.Resize(256),
            T.CenterCrop(224),
            T.ToTensor(),
            T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ])

    @classmethod
    def INPUT_TYPES(cls):
        return {
            "required": {
                "input_image": ("IMAGE",),
            },
        }

    RETURN_TYPES = ("IMAGE", "MASK")
    FUNCTION = "detect_furniture"

    CATEGORY = "custom"

    def detect_furniture(self, input_image):
        input_image = Image.fromarray((input_image * 255).astype('uint8'))
        input_tensor = self.transforms(input_image).unsqueeze(0)
        with torch.no_grad():
            output = self.model(input_tensor)['out'][0]
        output_predictions = output.argmax(0)

        non_furniture_classes = list(range(1, 151))  # Adjust the range based on ADE20K classes
        furniture_classes = [5, 10, 20, 25, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71]  # Based on ADE20K
        non_furniture_classes = [cls for cls in non_furniture_classes if cls not in furniture_classes]

        mask = torch.zeros_like(output_predictions, dtype=torch.bool)
        for cls in non_furniture_classes:
            mask |= (output_predictions == cls)

        mask = ~mask
        masked_image = input_image * mask.unsqueeze(-1).float()

        return masked_image, mask

NODE_CLASS_MAPPINGS = {
    "FurnitureMask": FurnitureMaskNode
}