models.py

"""This file contains methods for inference and image generation."""
import logging
from typing import List, Tuple, Dict

import streamlit as st
import torch
import gc
import time
import numpy as np
from PIL import Image
from time import perf_counter
from contextlib import contextmanager
from scipy.signal import fftconvolve
from PIL import ImageFilter

from transformers import AutoImageProcessor, UperNetForSemanticSegmentation
from diffusers import ControlNetModel, UniPCMultistepScheduler
from diffusers import StableDiffusionInpaintPipeline

from config import WIDTH, HEIGHT
from palette import ade_palette
from stable_diffusion_controlnet_inpaint_img2img import StableDiffusionControlNetInpaintImg2ImgPipeline

LOGGING = logging.getLogger(__name__)

def flush():
    gc.collect()
    torch.cuda.empty_cache()

class ControlNetPipeline:
    def __init__(self):
        print(torch.__version__)
        self.in_use = False
        self.controlnet = ControlNetModel.from_pretrained(
        "BertChristiaens/controlnet-seg-room", torch_dtype=torch.float16)

        self.pipe = StableDiffusionControlNetInpaintImg2ImgPipeline.from_pretrained(
            "runwayml/stable-diffusion-inpainting",
            controlnet=self.controlnet,
            safety_checker=None,
            torch_dtype=torch.float16
        )

        self.pipe.scheduler = UniPCMultistepScheduler.from_config(self.pipe.scheduler.config)
        self.pipe.enable_xformers_memory_efficient_attention()
        # self.pipe.enable_attention_slicing("max")
        self.pipe = self.pipe.to("cuda")
        
        self.waiting_queue = []
        self.count = 0
    
    @property
    def queue_size(self):
        return len(self.waiting_queue)
    
    def __call__(self, **kwargs):
        self.count += 1
        number = self.count

        self.waiting_queue.append(number)
        
        # wait until the next number in the queue is the current number
        while self.waiting_queue[0] != number:
            print(f"Wait for your turn {number} in queue {self.waiting_queue}")
            time.sleep(0.5)
            pass

        # it's your turn, so remove the number from the queue
        # and call the function
        print("It's the turn of", self.count)
        results = self.pipe(**kwargs)
        self.waiting_queue.pop(0)
        flush()
        return results


@contextmanager
def catchtime(message: str) -> float:
    """Context manager to measure time
    Args:
        message (str): message to log
    Returns:
        float: time in seconds
    Yields:
        Iterator[float]: time in seconds
    """
    start = perf_counter()
    yield lambda: perf_counter() - start
    LOGGING.info('%s: %.3f seconds', message, perf_counter() - start)


def convolution(mask: Image.Image, size=9) -> Image:
    """Method to blur the mask
    Args:
        mask (Image): masking image
        size (int, optional): size of the blur. Defaults to 9.
    Returns:
        Image: blurred mask
    """
    mask = np.array(mask.convert("L"))
    conv = np.ones((size, size)) / size**2
    mask_blended = fftconvolve(mask, conv, 'same')
    mask_blended = mask_blended.astype(np.uint8).copy()

    border = size

    # replace borders with original values
    mask_blended[:border, :] = mask[:border, :]
    mask_blended[-border:, :] = mask[-border:, :]
    mask_blended[:, :border] = mask[:, :border]
    mask_blended[:, -border:] = mask[:, -border:]

    return Image.fromarray(mask_blended).convert("L")


def postprocess_image_masking(inpainted: Image, image: Image, mask: Image) -> Image:
    """Method to postprocess the inpainted image
    Args:
        inpainted (Image): inpainted image
        image (Image): original image
        mask (Image): mask
    Returns:
        Image: inpainted image
    """
    final_inpainted = Image.composite(inpainted.convert("RGBA"), image.convert("RGBA"), mask)
    return final_inpainted.convert("RGB")


@st.experimental_singleton(max_entries=5)
def get_controlnet() -> ControlNetModel:
    """Method to load the controlnet model
    Returns:
        ControlNetModel: controlnet model
    """
    pipe = ControlNetPipeline()
    return pipe


@st.experimental_singleton(max_entries=5)
def get_segmentation_pipeline() -> Tuple[AutoImageProcessor, UperNetForSemanticSegmentation]:
    """Method to load the segmentation pipeline
    Returns:
        Tuple[AutoImageProcessor, UperNetForSemanticSegmentation]: segmentation pipeline
    """
    image_processor = AutoImageProcessor.from_pretrained("openmmlab/upernet-convnext-small")
    image_segmentor = UperNetForSemanticSegmentation.from_pretrained(
        "openmmlab/upernet-convnext-small")
    return image_processor, image_segmentor


@st.experimental_singleton(max_entries=5)
def get_inpainting_pipeline() -> StableDiffusionInpaintPipeline:
    """Method to load the inpainting pipeline
    Returns:
        StableDiffusionInpaintPipeline: inpainting pipeline
    """
    pipe = StableDiffusionInpaintPipeline.from_pretrained(
        "stabilityai/stable-diffusion-2-inpainting",
        torch_dtype=torch.float16,
        safety_checker=None,
    )

    pipe.enable_xformers_memory_efficient_attention()
    pipe = pipe.to("cuda")

    return pipe


def make_grid_parameters(grid_search: Dict, params: Dict) -> List[Dict]:
    """Method to make grid parameters
    Args:
        grid_search (Dict): grid search parameters
        params (Dict): fixed parameters
    Returns:
        List[Dict]: grid parameters
    """
    options = []

    for k in range(len(grid_search['generator'])):
        for i in range(len(grid_search['strength'])):
            for j in range(len(grid_search['guidance_scale'])):
                options.append({'strength': grid_search['strength'][i],
                                'guidance_scale': grid_search['guidance_scale'][j],
                                'generator': grid_search['generator'][k],
                                **params
                                })
    return options


def make_captions(options: List[Dict]) -> List[str]:
    """Method to make captions
    Args:
        options (List[Dict]): grid parameters
    Returns:
        List[str]: captions
    """
    captions = []
    for option in options:
        captions.append(
            f"strength {option['strength']}, guidance {option['guidance_scale']}, steps {option['num_inference_steps']}")
    return captions


@torch.inference_mode()
def make_image_controlnet(image: np.ndarray,
                          mask_image: np.ndarray,
                          controlnet_conditioning_image: np.ndarray,
                          positive_prompt: str, negative_prompt: str,
                          seed: int = 2356132) -> List[Image.Image]:
    """Method to make image using controlnet
    Args:
        image (np.ndarray): input image
        mask_image (np.ndarray): mask image
        controlnet_conditioning_image (np.ndarray): conditioning image
        positive_prompt (str): positive prompt string
        negative_prompt (str): negative prompt string
        seed (int, optional): seed. Defaults to 2356132.
    Returns:
        List[Image.Image]: list of generated images
    """

    with catchtime("get controlnet"):
        pipe = get_controlnet()

    torch.cuda.empty_cache()
    images = []

    common_parameters = {'prompt': positive_prompt,
                        'negative_prompt': negative_prompt,
                        'num_inference_steps': 30,
                            'controlnet_conditioning_scale': 1.1,
                            'controlnet_conditioning_scale_decay': 0.96,
                            'controlnet_steps': 28,
                        }

    grid_search = {'strength': [1.00, ],
                   'guidance_scale': [7.0],
                   'generator': [[torch.Generator(device="cuda").manual_seed(seed+i)] for i in range(1)],
                   }

    prompt_settings = make_grid_parameters(grid_search, common_parameters)


    mask_image = Image.fromarray((mask_image * 255).astype(np.uint8)).convert("RGB")
    image = Image.fromarray(image).convert("RGB")
    controlnet_conditioning_image = Image.fromarray(controlnet_conditioning_image).convert("RGB").filter(ImageFilter.GaussianBlur(radius = 9))

    mask_image_postproc = convolution(mask_image)

    with catchtime("Controlnet generation total"):
        for _, setting in enumerate(prompt_settings):
            st.success(f"{pipe.queue_size} images in the queue, can take up to {pipe.queue_size * 20} seconds")
            with catchtime("Controlnet generation"):
                generated_image = pipe(
                    **setting,
                    image=image,
                    mask_image=mask_image,
                    controlnet_conditioning_image=controlnet_conditioning_image,
                ).images[0]
                generated_image = postprocess_image_masking(
                    generated_image, image, mask_image_postproc)
            images.append(generated_image)

    return images


@torch.inference_mode()
def make_inpainting(positive_prompt: str,
                    image: Image,
                    mask_image: np.ndarray,
                    negative_prompt: str = "") -> List[Image.Image]:
    """Method to make inpainting
    Args:
        positive_prompt (str): positive prompt string
        image (Image): input image
        mask_image (np.ndarray): mask image
        negative_prompt (str, optional): negative prompt string. Defaults to "".
    Returns:
        List[Image.Image]: list of generated images
    """

    with catchtime("Get inpainting pipeline"):
        pipe = get_inpainting_pipeline()

    common_parameters = {'prompt': positive_prompt,
                        'negative_prompt': negative_prompt,
                        'num_inference_steps': 20,
                        }

    torch.cuda.empty_cache()
    images = []
    for _ in range(1):
        with catchtime("Inpainting generation"):
            image_ = pipe(image=image,
                          mask_image=Image.fromarray((mask_image * 255).astype(np.uint8)),
                          height=HEIGHT,
                          width=WIDTH,
                          **common_parameters
                          ).images[0]
        images.append(image_)
    return images


@torch.inference_mode()
@torch.autocast('cuda')
def segment_image(image: Image) -> Image:
    """Method to segment image
    Args:
        image (Image): input image
    Returns:
        Image: segmented image
    """
    image_processor, image_segmentor = get_segmentation_pipeline()
    pixel_values = image_processor(image, return_tensors="pt").pixel_values
    with torch.no_grad():
        outputs = image_segmentor(pixel_values)

    seg = image_processor.post_process_semantic_segmentation(
        outputs, target_sizes=[image.size[::-1]])
    seg = seg[0]
    color_seg = np.zeros((seg.shape[0], seg.shape[1], 3), dtype=np.uint8)  # height, width, 3
    palette = np.array(ade_palette())
    for label, color in enumerate(palette):
        color_seg[seg == label, :] = color
    color_seg = color_seg.astype(np.uint8)
    seg_image = Image.fromarray(color_seg).convert('RGB')
    return seg_image