grounded_sam/app.py

import os, sys
import random
import time
import warnings
from scipy.ndimage import binary_dilation
from PIL import Image, ImageDraw, ImageFont

def downloadStuff():
    # os.system("python -m pip install -e segment_anything")
    # os.system("python -m pip install -e GroundingDINO")
    os.system("pip install --upgrade diffusers[torch]")
    # os.system("pip install opencv-python pycocotools matplotlib onnxruntime onnx ipykernel")
    # os.system("wget https://github.com/IDEA-Research/Grounded-Segment-Anything/raw/main/assets/demo1.jpg")
    # os.system("wget https://huggingface.co/ShilongLiu/GroundingDINO/resolve/main/groundingdino_swint_ogc.pth")
    # os.system("wget https://huggingface.co/spaces/mrtlive/segment-anything-model/resolve/main/sam_vit_h_4b8939.pth")
    # os.system("wget https://huggingface.co/lkeab/hq-sam/resolve/main/sam_hq_vit_h.pth")
    sys.path.append(os.path.join(os.getcwd(), "GroundingDINO"))
    sys.path.append(os.path.join(os.getcwd(), "segment_anything"))

# downloadStuff()

warnings.filterwarnings("ignore")

import gradio as gr
import argparse

import numpy as np
import torch
import torchvision
from PIL import Image, ImageDraw, ImageFont

# Grounding DINO
from .GroundingDINO.groundingdino.datasets import transforms as T
from .GroundingDINO.groundingdino.models import build_model
from .GroundingDINO.groundingdino.util.slconfig import SLConfig
from .GroundingDINO.groundingdino.util.utils import clean_state_dict, get_phrases_from_posmap

# segment anything
# from .segment_anything.segment_anything import build_sam, SamPredictor
from .segment_anything.segment_anything.build_sam import build_sam
from .segment_anything.segment_anything.predictor import SamPredictor
import numpy as np

# diffusers
import torch
from diffusers import StableDiffusionInpaintPipeline

# BLIP
from transformers import BlipProcessor, BlipForConditionalGeneration


def generate_caption(processor, blip_model, raw_image):
    # unconditional image captioning
    inputs = processor(raw_image, return_tensors="pt").to(
        "cuda", torch.float16)
    out = blip_model.generate(**inputs)
    caption = processor.decode(out[0], skip_special_tokens=True)
    return caption


def transform_image(image_pil):

    transform = T.Compose(
        [
            T.RandomResize([800], max_size=1333),
            T.ToTensor(),
            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ]
    )
    image, _ = transform(image_pil, None)  # 3, h, w
    return image


def load_model(model_config_path, model_checkpoint_path, device):
    args = SLConfig.fromfile(model_config_path)
    args.device = device
    model = build_model(args)
    checkpoint = torch.load(model_checkpoint_path, map_location="cpu")
    load_res = model.load_state_dict(
        clean_state_dict(checkpoint["model"]), strict=False)
    print(load_res)
    _ = model.eval()
    return model


def get_grounding_output(model, image, caption, box_threshold, text_threshold, with_logits=True):
    image = image.to("cuda")
    
    caption = caption.lower()
    caption = caption.strip()
    if not caption.endswith("."):
        caption = caption + "."

    start_time = time.time()
    model.to("cuda") 
    with torch.no_grad():
        outputs = model(image[None], captions=[caption])
    print(f"Model forward time: {time.time() - start_time}")

    start_time = time.time()
    logits = outputs["pred_logits"].cpu().sigmoid()[0]  # (nq, 256)
    boxes = outputs["pred_boxes"].cpu()[0]  # (nq, 4)
    logits.shape[0]
    print(f"Process output time: {time.time() - start_time}")

    # filter output
    start_time = time.time()
    logits_filt = logits.clone()
    boxes_filt = boxes.clone()
    filt_mask = logits_filt.max(dim=1)[0] > box_threshold
    logits_filt = logits_filt[filt_mask]  # num_filt, 256
    boxes_filt = boxes_filt[filt_mask]  # num_filt, 4
    logits_filt.shape[0]
    print(f"Filter output time: {time.time() - start_time}")

    # get phrase
    start_time = time.time()
    tokenlizer = model.tokenizer
    tokenized = tokenlizer(caption)
    print(f"Tokenize time: {time.time() - start_time}")
    # build pred
    pred_phrases = []
    scores = []
    start_time = time.time()
    for logit, box in zip(logits_filt, boxes_filt):
        pred_phrase = get_phrases_from_posmap(
            logit > text_threshold, tokenized, tokenlizer)
        if with_logits:
            pred_phrases.append(
                pred_phrase + f"({str(logit.max().item())[:4]})")
        else:
            pred_phrases.append(pred_phrase)
        scores.append(logit.max().item())

    print(f"Build pred time: {time.time() - start_time}")

    return boxes_filt, torch.Tensor(scores), pred_phrases


def draw_mask(mask, draw, random_color=False):
    if random_color:
        color = (random.randint(0, 255), random.randint(
            0, 255), random.randint(0, 255), 153)
    else:
        color = (30, 144, 255, 153)

    nonzero_coords = np.transpose(np.nonzero(mask))

    for coord in nonzero_coords:
        draw.point(coord[::-1], fill=color)


def draw_box(box, draw, label):
    # random color
    color = tuple(np.random.randint(0, 255, size=3).tolist())

    draw.rectangle(((box[0], box[1]), (box[2], box[3])),
                   outline=color,  width=2)

    if label:
        font = ImageFont.load_default()
        if hasattr(font, "getbbox"):
            bbox = draw.textbbox((box[0], box[1]), str(label), font)
        else:
            w, h = draw.textsize(str(label), font)
            bbox = (box[0], box[1], w + box[0], box[1] + h)
        draw.rectangle(bbox, fill=color)
        draw.text((box[0], box[1]), str(label), fill="white")

        draw.text((box[0], box[1]), label)


config_file = f'{os.environ.get("path", ".")}/grounded_sam/GroundingDINO/groundingdino/config/GroundingDINO_SwinT_OGC.py'
ckpt_repo_id = "ShilongLiu/GroundingDINO"
ckpt_filenmae = f'{os.environ.get("path", ".")}/checkpoints/groundingdino_swint_ogc.pth'
sam_checkpoint = f'{os.environ.get("path", ".")}/checkpoints/sam_hq_vit_h.pth'
output_dir = "outputs"
device = 'cuda' if torch.cuda.is_available() else 'cpu'


blip_processor = None
blip_model = None
groundingdino_model = load_model(
            config_file, ckpt_filenmae, device=device)
sam_predictor = None
inpaint_pipeline = None


def run_grounded_sam(input_image, text_prompt, task_type, inpaint_prompt, box_threshold, text_threshold, iou_threshold, inpaint_mode):

    global blip_processor, blip_model, groundingdino_model, sam_predictor, inpaint_pipeline

    # make dir
    os.makedirs(output_dir, exist_ok=True)
    # load image
    start_time = time.time()
    image_pil = input_image.convert("RGB")
    transformed_image = transform_image(image_pil)
    print(f"Transform image time: {time.time() - start_time}")

    start_time = time.time()
    print(f"Load model time: {time.time() - start_time}")

    # run grounding dino model
    start_time = time.time()
    boxes_filt, scores, pred_phrases = get_grounding_output(
        groundingdino_model, transformed_image, text_prompt, box_threshold, text_threshold
    )
    print(f"Run model time: {time.time() - start_time}")

    size = image_pil.size

    # process boxes
    start_time = time.time()
    H, W = size[1], size[0]
    for i in range(boxes_filt.size(0)):
        boxes_filt[i] = boxes_filt[i] * torch.Tensor([W, H, W, H])
        boxes_filt[i][:2] -= boxes_filt[i][2:] / 2
        boxes_filt[i][2:] += boxes_filt[i][:2]

    boxes_filt = boxes_filt.cpu()
    print(f"Process boxes time: {time.time() - start_time}")

    # nms
    start_time = time.time()
    print(f"Before NMS: {boxes_filt.shape[0]} boxes")
    nms_idx = torchvision.ops.nms(
        boxes_filt, scores, iou_threshold).numpy().tolist()
    boxes_filt = boxes_filt[nms_idx]
    pred_phrases = [pred_phrases[idx] for idx in nms_idx]
    print(f"After NMS: {boxes_filt.shape[0]} boxes")
    print(f"NMS time: {time.time() - start_time}")

    start_time = time.time()
    if sam_predictor is None:
        # initialize SAM
        assert sam_checkpoint, 'sam_checkpoint is not found!'
        sam = build_sam(checkpoint=sam_checkpoint)
        sam.to(device=device)
        sam_predictor = SamPredictor(sam)
    print(f"Initialize SAM time: {time.time() - start_time}")

    image = np.array(image_pil)
    sam_predictor.set_image(image)

    start_time = time.time()
    transformed_boxes = sam_predictor.transform.apply_boxes_torch(
        boxes_filt, image.shape[:2]).to(device)
    print(f"Transform boxes time: {time.time() - start_time}")

    start_time = time.time()
    masks, _, _ = sam_predictor.predict_torch(
        point_coords=None,
        point_labels=None,
        boxes=transformed_boxes,
        multimask_output=False,
    )
    print(f"Predict time: {time.time() - start_time}")

    # masks: [1, 1, 512, 512]


    mask_image = Image.new('RGBA', size, color=(0, 0, 0, 0))

    mask_draw = ImageDraw.Draw(mask_image)
    for mask in masks:
        draw_mask(mask[0].cpu().numpy(), mask_draw, random_color=True)

    image_draw = ImageDraw.Draw(image_pil)

    for box, label in zip(boxes_filt, pred_phrases):
        draw_box(box, image_draw, label)

    image_pil = image_pil.convert('RGBA')
    image_pil.alpha_composite(mask_image)

    return [image_pil, mask_image]

def toBlackWhiteMask(input_pil):
    if input_pil.mode != 'RGBA':
        input_pil = input_pil.convert('RGBA')

    new_image = Image.new('RGBA', input_pil.size, (255, 255, 255, 255))

    input_data = input_pil.getdata()
    new_data = []

    for item in input_data:
        if item[3] == 0:
            new_data.append((0, 0, 0, 255))
        else:
            new_data.append((255, 255, 255, 255))

    new_image.putdata(new_data)

    return new_image

def expand_white_pixels(input_pil, expand_by=1):
    # Convert the input image to grayscale
    grayscale = input_pil.convert('L')

    # Create a binary mask where white pixels are represented by 1
    binary_mask = np.array(grayscale) > 245

    # Apply the dilation operation to the binary mask
    dilated_mask = binary_dilation(binary_mask, iterations=expand_by)

    # Create a new PIL image from the dilated mask
    expanded_image = Image.fromarray(np.uint8(dilated_mask * 255))

    return expanded_image

def get_mask(input_pil, positive_prompt, expand_by=0):
    result = run_grounded_sam(input_image=input_pil, text_prompt=positive_prompt, task_type="seg", inpaint_prompt=None, box_threshold=0.3,
     text_threshold=0.25, iou_threshold=0.8, inpaint_mode="merge")

    result = result[1]

    black_white_result = toBlackWhiteMask(result)

    if expand_by > 0:
        black_white_result = expand_white_pixels(black_white_result, expand_by=expand_by)

    return black_white_result

if __name__ == "__main__":
    parser = argparse.ArgumentParser("Grounded SAM demo", add_help=True)
    parser.add_argument("--debug", action="store_true",
                        help="using debug mode")
    parser.add_argument("--share", action="store_true", help="share the app")
    parser.add_argument('--no-gradio-queue', action="store_true",
                        help='path to the SAM checkpoint')
    args = parser.parse_args()

    print(args)

    block = gr.Blocks()
    if not args.no_gradio_queue:
        block = block.queue()

    with block:
        with gr.Row():
            with gr.Column():
                input_image = gr.Image(
                    source='upload', type="pil", value="demo1.jpg")
                task_type = gr.Dropdown(
                    ["det", "seg", "inpainting", "automatic"], value="seg", label="task_type")
                text_prompt = gr.Textbox(label="Text Prompt", placeholder="bear . beach .")
                inpaint_prompt = gr.Textbox(label="Inpaint Prompt", placeholder="A dinosaur, detailed, 4K.")
                run_button = gr.Button(label="Run")
                with gr.Accordion("Advanced options", open=False):
                    box_threshold = gr.Slider(
                        label="Box Threshold", minimum=0.0, maximum=1.0, value=0.3, step=0.001
                    )
                    text_threshold = gr.Slider(
                        label="Text Threshold", minimum=0.0, maximum=1.0, value=0.25, step=0.001
                    )
                    iou_threshold = gr.Slider(
                        label="IOU Threshold", minimum=0.0, maximum=1.0, value=0.8, step=0.001
                    )
                    inpaint_mode = gr.Dropdown(
                        ["merge", "first"], value="merge", label="inpaint_mode")

            with gr.Column():
                gallery = gr.Gallery(
                    label="Generated images", show_label=False, elem_id="gallery"
                ).style(preview=True, grid=2, object_fit="scale-down")

        run_button.click(fn=run_grounded_sam, inputs=[
            input_image, text_prompt, task_type, inpaint_prompt, box_threshold, text_threshold, iou_threshold, inpaint_mode], outputs=gallery)

    block.launch(debug=args.debug, share=args.share, show_error=True)