import warnings warnings.filterwarnings('ignore') import subprocess, io, os, sys, time from loguru import logger os.environ["CUDA_VISIBLE_DEVICES"] = "0" if os.environ.get('IS_MY_DEBUG') is None: result = subprocess.run(['pip', 'install', '-e', 'GroundingDINO'], check=True) print(f'pip install GroundingDINO = {result}') result = subprocess.run(['pip', 'list'], check=True) print(f'pip list = {result}') sys.path.insert(0, './GroundingDINO') if not os.path.exists('./sam_vit_h_4b8939.pth'): logger.info(f"get sam_vit_h_4b8939.pth...") result = subprocess.run(['wget', 'https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth'], check=True) print(f'wget sam_vit_h_4b8939.pth result = {result}') import gradio as gr import argparse import copy import numpy as np import torch from PIL import Image, ImageDraw, ImageFont, ImageOps # Grounding DINO import GroundingDINO.groundingdino.datasets.transforms as T from GroundingDINO.groundingdino.models import build_model from GroundingDINO.groundingdino.util import box_ops from GroundingDINO.groundingdino.util.slconfig import SLConfig from GroundingDINO.groundingdino.util.utils import clean_state_dict, get_phrases_from_posmap import cv2 import numpy as np import matplotlib.pyplot as plt from lama_cleaner.model_manager import ModelManager from lama_cleaner.schema import Config # segment anything from segment_anything import build_sam, SamPredictor # diffusers import PIL import requests import torch from io import BytesIO from diffusers import StableDiffusionInpaintPipeline from huggingface_hub import hf_hub_download def get_device(): from numba import cuda if cuda.is_available(): device = 'cuda:0' # cuda.get_current_device() else: device = 'cpu' return device def load_model_hf(model_config_path, repo_id, filename, device='cpu'): args = SLConfig.fromfile(model_config_path) model = build_model(args) args.device = device cache_file = hf_hub_download(repo_id=repo_id, filename=filename) checkpoint = torch.load(cache_file, map_location=device) log = model.load_state_dict(clean_state_dict(checkpoint['model']), strict=False) print("Model loaded from {} \n => {}".format(cache_file, log)) _ = model.eval() return model def plot_boxes_to_image(image_pil, tgt): H, W = tgt["size"] boxes = tgt["boxes"] labels = tgt["labels"] assert len(boxes) == len(labels), "boxes and labels must have same length" draw = ImageDraw.Draw(image_pil) mask = Image.new("L", image_pil.size, 0) mask_draw = ImageDraw.Draw(mask) # draw boxes and masks for box, label in zip(boxes, labels): # from 0..1 to 0..W, 0..H box = box * torch.Tensor([W, H, W, H]) # from xywh to xyxy box[:2] -= box[2:] / 2 box[2:] += box[:2] # random color color = tuple(np.random.randint(0, 255, size=3).tolist()) # draw x0, y0, x1, y1 = box x0, y0, x1, y1 = int(x0), int(y0), int(x1), int(y1) draw.rectangle([x0, y0, x1, y1], outline=color, width=6) # draw.text((x0, y0), str(label), fill=color) font = ImageFont.load_default() if hasattr(font, "getbbox"): bbox = draw.textbbox((x0, y0), str(label), font) else: w, h = draw.textsize(str(label), font) bbox = (x0, y0, w + x0, y0 + h) # bbox = draw.textbbox((x0, y0), str(label)) draw.rectangle(bbox, fill=color) draw.text((x0, y0), str(label), fill="white") mask_draw.rectangle([x0, y0, x1, y1], fill=255, width=6) return image_pil, mask def load_image(image_path): # # load image if isinstance(image_path, PIL.Image.Image): image_pil = image_path else: image_pil = Image.open(image_path).convert("RGB") # load image 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_pil, 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=device) #"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, device="cpu"): caption = caption.lower() caption = caption.strip() if not caption.endswith("."): caption = caption + "." model = model.to(device) image = image.to(device) with torch.no_grad(): outputs = model(image[None], captions=[caption]) logits = outputs["pred_logits"].cpu().sigmoid()[0] # (nq, 256) boxes = outputs["pred_boxes"].cpu()[0] # (nq, 4) logits.shape[0] # filter output 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] # get phrase tokenlizer = model.tokenizer tokenized = tokenlizer(caption) # build pred pred_phrases = [] 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) return boxes_filt, pred_phrases def show_mask(mask, ax, random_color=False): if random_color: color = np.concatenate([np.random.random(3), np.array([0.6])], axis=0) else: color = np.array([30/255, 144/255, 255/255, 0.6]) h, w = mask.shape[-2:] mask_image = mask.reshape(h, w, 1) * color.reshape(1, 1, -1) ax.imshow(mask_image) def show_box(box, ax, label): x0, y0 = box[0], box[1] w, h = box[2] - box[0], box[3] - box[1] ax.add_patch(plt.Rectangle((x0, y0), w, h, edgecolor='green', facecolor=(0,0,0,0), lw=2)) ax.text(x0, y0, label) def xywh_to_xyxy(box, sizeW, sizeH): if isinstance(box, list): box = torch.Tensor(box) box = box * torch.Tensor([sizeW, sizeH, sizeW, sizeH]) box[:2] -= box[2:] / 2 box[2:] += box[:2] box = box.numpy() return box def mask_extend(img, box, extend_pixels=10, useRectangle=True): box[0] = int(box[0]) box[1] = int(box[1]) box[2] = int(box[2]) box[3] = int(box[3]) region = img.crop(tuple(box)) new_width = box[2] - box[0] + 2*extend_pixels new_height = box[3] - box[1] + 2*extend_pixels region_BILINEAR = region.resize((int(new_width), int(new_height))) if useRectangle: region_draw = ImageDraw.Draw(region_BILINEAR) region_draw.rectangle((0, 0, new_width, new_height), fill=(255, 255, 255)) img.paste(region_BILINEAR, (int(box[0]-extend_pixels), int(box[1]-extend_pixels))) return img def mix_masks(imgs): re_img = 1 - np.asarray(imgs[0].convert("1")) for i in range(len(imgs)-1): re_img = np.multiply(re_img, 1 - np.asarray(imgs[i+1].convert("1"))) re_img = 1 - re_img return Image.fromarray(np.uint8(255*re_img)) config_file = 'GroundingDINO/groundingdino/config/GroundingDINO_SwinT_OGC.py' ckpt_repo_id = "ShilongLiu/GroundingDINO" ckpt_filenmae = "groundingdino_swint_ogc.pth" sam_checkpoint = './sam_vit_h_4b8939.pth' output_dir = "outputs" device = "cuda" device = get_device() print(f'device={device}') # make dir os.makedirs(output_dir, exist_ok=True) # initialize groundingdino model logger.info(f"initialize groundingdino model...") groundingdino_model = load_model_hf(config_file, ckpt_repo_id, ckpt_filenmae) # initialize SAM logger.info(f"initialize SAM model...") sam_predictor = SamPredictor(build_sam(checkpoint=sam_checkpoint)) # initialize stable-diffusion-inpainting logger.info(f"initialize stable-diffusion-inpainting...") sd_pipe = None if os.environ.get('IS_MY_DEBUG') is None: sd_pipe = StableDiffusionInpaintPipeline.from_pretrained( "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16 ) sd_pipe = sd_pipe.to(device) # initialize lama_cleaner logger.info(f"initialize lama_cleaner...") from lama_cleaner.helper import ( load_img, numpy_to_bytes, resize_max_size, ) lama_cleaner_model = ModelManager( name='lama', device='cpu', # device, ) def lama_cleaner_process(image, mask): ori_image = image if mask.shape[0] == image.shape[1] and mask.shape[1] == image.shape[0] and mask.shape[0] != mask.shape[1]: # rotate image ori_image = np.transpose(image[::-1, ...][:, ::-1], axes=(1, 0, 2))[::-1, ...] image = ori_image original_shape = ori_image.shape interpolation = cv2.INTER_CUBIC size_limit = 1080 if size_limit == "Original": size_limit = max(image.shape) else: size_limit = int(size_limit) config = Config( ldm_steps=25, ldm_sampler='plms', zits_wireframe=True, hd_strategy='Original', hd_strategy_crop_margin=196, hd_strategy_crop_trigger_size=1280, hd_strategy_resize_limit=2048, prompt='', use_croper=False, croper_x=0, croper_y=0, croper_height=512, croper_width=512, sd_mask_blur=5, sd_strength=0.75, sd_steps=50, sd_guidance_scale=7.5, sd_sampler='ddim', sd_seed=42, cv2_flag='INPAINT_NS', cv2_radius=5, ) if config.sd_seed == -1: config.sd_seed = random.randint(1, 999999999) # logger.info(f"Origin image shape_0_: {original_shape} / {size_limit}") image = resize_max_size(image, size_limit=size_limit, interpolation=interpolation) # logger.info(f"Resized image shape_1_: {image.shape}") # logger.info(f"mask image shape_0_: {mask.shape} / {type(mask)}") mask = resize_max_size(mask, size_limit=size_limit, interpolation=interpolation) # logger.info(f"mask image shape_1_: {mask.shape} / {type(mask)}") res_np_img = lama_cleaner_model(image, mask, config) torch.cuda.empty_cache() image = Image.open(io.BytesIO(numpy_to_bytes(res_np_img, 'png'))) return image mask_source_draw = "draw a mask on input image" mask_source_segment = "type what to detect below" def run_grounded_sam(input_image, text_prompt, task_type, inpaint_prompt, box_threshold, text_threshold, iou_threshold, inpaint_mode, mask_source_radio, remove_mode, remove_mask_extend): text_prompt = text_prompt.strip() if not ((task_type == 'inpainting' or task_type == 'remove') and mask_source_radio == mask_source_draw): if text_prompt == '': return [], gr.Gallery.update(label='Detection prompt is not found!😂😂😂😂') if input_image is None: return [], gr.Gallery.update(label='Please upload a image!😂😂😂😂') file_temp = int(time.time()) logger.info(f'run_grounded_sam_[{file_temp}]_{task_type}/{inpaint_mode}/[{mask_source_radio}]/{remove_mode}/{remove_mask_extend}_[{text_prompt}]/[{inpaint_prompt}]___1_') # load image input_mask_pil = input_image['mask'] input_mask = np.array(input_mask_pil.convert("L")) image_pil, image = load_image(input_image['image'].convert("RGB")) # visualize raw image # image_pil.save(os.path.join(output_dir, f"raw_image_{file_temp}.jpg")) size = image_pil.size output_images = [] # run grounding dino model if (task_type == 'inpainting' or task_type == 'remove') and mask_source_radio == mask_source_draw: pass else: groundingdino_device = 'cpu' if device != 'cpu': try: from groundingdino import _C groundingdino_device = 'cuda:0' except: warnings.warn("Failed to load custom C++ ops. Running on CPU mode Only in groundingdino!") groundingdino_device = 'cpu' boxes_filt, pred_phrases = get_grounding_output( groundingdino_model, image, text_prompt, box_threshold, text_threshold, device=groundingdino_device ) if boxes_filt.size(0) == 0: logger.info(f'run_grounded_sam_[{file_temp}]_{task_type}_[{text_prompt}]_1_[No objects detected, please try others.]_') return [], gr.Gallery.update(label='No objects detected, please try others.😂😂😂😂') boxes_filt_ori = copy.deepcopy(boxes_filt) pred_dict = { "boxes": boxes_filt, "size": [size[1], size[0]], # H,W "labels": pred_phrases, } image_with_box = plot_boxes_to_image(copy.deepcopy(image_pil), pred_dict)[0] image_path = os.path.join(output_dir, f"grounding_dino_output_{file_temp}.jpg") image_with_box.save(image_path) detection_image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB) os.remove(image_path) output_images.append(detection_image_result) logger.info(f'run_grounded_sam_[{file_temp}]_{task_type}_2_') if task_type == 'segment' or ((task_type == 'inpainting' or task_type == 'remove') and mask_source_radio == mask_source_segment): image = np.array(input_image['image']) sam_predictor.set_image(image) 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() transformed_boxes = sam_predictor.transform.apply_boxes_torch(boxes_filt, image.shape[:2]) masks, _, _ = sam_predictor.predict_torch( point_coords = None, point_labels = None, boxes = transformed_boxes, multimask_output = False, ) # masks: [9, 1, 512, 512] assert sam_checkpoint, 'sam_checkpoint is not found!' # draw output image plt.figure(figsize=(10, 10)) plt.imshow(image) for mask in masks: show_mask(mask.cpu().numpy(), plt.gca(), random_color=True) for box, label in zip(boxes_filt, pred_phrases): show_box(box.numpy(), plt.gca(), label) plt.axis('off') image_path = os.path.join(output_dir, f"grounding_seg_output_{file_temp}.jpg") plt.savefig(image_path, bbox_inches="tight") segment_image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB) os.remove(image_path) output_images.append(segment_image_result) logger.info(f'run_grounded_sam_[{file_temp}]_{task_type}_3_') if task_type == 'detection' or task_type == 'segment': logger.info(f'run_grounded_sam_[{file_temp}]_{task_type}_9_') return output_images, gr.Gallery.update(label='result images') elif task_type == 'inpainting' or task_type == 'remove': if inpaint_prompt.strip() == '' and mask_source_radio == mask_source_segment: task_type = 'remove' logger.info(f'run_grounded_sam_[{file_temp}]_{task_type}_4_') if mask_source_radio == mask_source_draw: mask_pil = input_mask_pil mask = input_mask else: masks_ori = copy.deepcopy(masks) if inpaint_mode == 'merge': masks = torch.sum(masks, dim=0).unsqueeze(0) masks = torch.where(masks > 0, True, False) mask = masks[0][0].cpu().numpy() mask_pil = Image.fromarray(mask) image_path = os.path.join(output_dir, f"image_mask_{file_temp}.jpg") mask_pil.convert("RGB").save(image_path) image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB) os.remove(image_path) output_images.append(image_result) if task_type == 'inpainting': # inpainting pipeline image_source_for_inpaint = image_pil.resize((512, 512)) image_mask_for_inpaint = mask_pil.resize((512, 512)) image_inpainting = sd_pipe(prompt=inpaint_prompt, image=image_source_for_inpaint, mask_image=image_mask_for_inpaint).images[0] else: # remove from mask if mask_source_radio == mask_source_segment: mask_imgs = [] masks_shape = masks_ori.shape boxes_filt_ori_array = boxes_filt_ori.numpy() if inpaint_mode == 'merge': extend_shape_0 = masks_shape[0] extend_shape_1 = masks_shape[1] else: extend_shape_0 = 1 extend_shape_1 = 1 for i in range(extend_shape_0): for j in range(extend_shape_1): mask = masks_ori[i][j].cpu().numpy() mask_pil = Image.fromarray(mask) if remove_mode == 'segment': useRectangle = False else: useRectangle = True try: remove_mask_extend = int(remove_mask_extend) except: remove_mask_extend = 10 mask_pil_exp = mask_extend(copy.deepcopy(mask_pil).convert("RGB"), xywh_to_xyxy(torch.tensor(boxes_filt_ori_array[i]), size[0], size[1]), extend_pixels=remove_mask_extend, useRectangle=useRectangle) mask_imgs.append(mask_pil_exp) mask_pil = mix_masks(mask_imgs) image_path = os.path.join(output_dir, f"image_mask_{file_temp}.jpg") mask_pil.convert("RGB").save(image_path) image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB) os.remove(image_path) output_images.append(image_result) image_inpainting = lama_cleaner_process(np.array(image_pil), np.array(mask_pil.convert("L"))) image_inpainting = image_inpainting.resize((image_pil.size[0], image_pil.size[1])) image_path = os.path.join(output_dir, f"grounded_sam_inpainting_output_{file_temp}.jpg") image_inpainting.save(image_path) image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB) os.remove(image_path) logger.info(f'run_grounded_sam_[{file_temp}]_{task_type}_9_') output_images.append(image_result) return output_images, gr.Gallery.update(label='result images') else: logger.info(f"task_type:{task_type} error!") logger.info(f'run_grounded_sam_[{file_temp}]_9_9_') return output_images, gr.Gallery.update(label='result images') def change_radio_display(task_type, mask_source_radio): text_prompt_visible = True inpaint_prompt_visible = False mask_source_radio_visible = False if task_type == "inpainting": inpaint_prompt_visible = True if task_type == "inpainting" or task_type == "remove": mask_source_radio_visible = True if mask_source_radio == mask_source_draw: text_prompt_visible = False return gr.Textbox.update(visible=text_prompt_visible), gr.Textbox.update(visible=inpaint_prompt_visible), gr.Radio.update(visible=mask_source_radio_visible) 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") args = parser.parse_args() print(f'args = {args}') block = gr.Blocks().queue() with block: with gr.Row(): with gr.Column(): input_image = gr.Image(source='upload', elem_id="image_upload", tool='sketch', type='pil', label="Upload") task_type = gr.Radio(["detection", "segment", "inpainting", "remove"], value="detection", label='Task type',interactive=True, visible=True) mask_source_radio = gr.Radio([mask_source_draw, mask_source_segment], value=mask_source_segment, label="Mask from", interactive=True, visible=False) text_prompt = gr.Textbox(label="Detection Prompt[To detect multiple objects, seperating each name with '.', like this: cat . dog . chair ]", placeholder="Cannot be empty") inpaint_prompt = gr.Textbox(label="Inpaint Prompt (if this is empty, then remove)", visible=False) 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=0.6, step=0.001 ) text_threshold = gr.Slider( label="Text Threshold", minimum=0.0, maximum=1.0, value=0.5, step=0.001 ) iou_threshold = gr.Slider( label="IOU Threshold", minimum=0.0, maximum=1.0, value=0.5, step=0.001 ) inpaint_mode = gr.Radio(["merge", "first"], value="merge", label="inpaint_mode") with gr.Row(): with gr.Column(scale=1): remove_mode = gr.Radio(["segment", "rectangle"], value="segment", label='remove mode') with gr.Column(scale=1): remove_mask_extend = gr.Textbox(label="remove_mask_extend", value='10') with gr.Column(): gallery = gr.Gallery( label="result images", show_label=True, elem_id="gallery" ).style(grid=[2], full_width=True, full_height=True) run_button.click(fn=run_grounded_sam, inputs=[ input_image, text_prompt, task_type, inpaint_prompt, box_threshold, text_threshold, iou_threshold, inpaint_mode, mask_source_radio, remove_mode, remove_mask_extend], outputs=[gallery, gallery]) task_type.change(fn=change_radio_display, inputs=[task_type, mask_source_radio], outputs=[text_prompt, inpaint_prompt, mask_source_radio]) mask_source_radio.change(fn=change_radio_display, inputs=[task_type, mask_source_radio], outputs=[text_prompt, inpaint_prompt, mask_source_radio]) DESCRIPTION = '### This demo from [Grounded-Segment-Anything](https://github.com/IDEA-Research/Grounded-Segment-Anything). Thanks for their excellent work.' DESCRIPTION += f'

For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings. Duplicate Space

' gr.Markdown(DESCRIPTION) block.launch(server_name='0.0.0.0', debug=args.debug, share=args.share)