|
import string |
|
import warnings |
|
warnings.filterwarnings('ignore') |
|
from entklei import get_nude |
|
from scipy.ndimage import binary_dilation |
|
|
|
import subprocess, io, os, sys, time |
|
|
|
is_production = False |
|
install_stuff = False |
|
os.environ['CUDA_HOME'] = '/usr/local/cuda-11.7/' if is_production else '/usr/local/cuda-12.1/' |
|
|
|
run_gradio = False |
|
|
|
if run_gradio and install_stuff: |
|
os.system("pip install gradio==3.50.2") |
|
|
|
import gradio as gr |
|
|
|
from loguru import logger |
|
|
|
os.environ["CUDA_VISIBLE_DEVICES"] = "0" |
|
|
|
if is_production: |
|
os.chdir("/repository") |
|
sys.path.insert(0, '/repository') |
|
|
|
if install_stuff: |
|
result = subprocess.run(['pip', 'install', '-e', 'GroundingDINO'], check=True) |
|
print(f'pip install GroundingDINO = {result}') |
|
|
|
|
|
|
|
|
|
sys.path.insert(0, '/repository/GroundingDINO' if is_production else "./GroundingDINO") |
|
|
|
import argparse |
|
import copy |
|
|
|
import numpy as np |
|
import torch |
|
from PIL import Image, ImageDraw, ImageFont, ImageOps |
|
|
|
|
|
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 |
|
matplotlib.use('AGG') |
|
plt = matplotlib.pyplot |
|
|
|
|
|
|
|
|
|
|
|
|
|
groundingdino_enable = True |
|
sam_enable = True |
|
inpainting_enable = True |
|
ram_enable = True |
|
|
|
lama_cleaner_enable = True |
|
|
|
kosmos_enable = False |
|
|
|
|
|
|
|
|
|
if os.environ.get('IS_MY_DEBUG') is not None: |
|
sam_enable = False |
|
ram_enable = False |
|
inpainting_enable = False |
|
kosmos_enable = False |
|
|
|
if lama_cleaner_enable: |
|
try: |
|
from lama_cleaner.model_manager import ModelManager |
|
from lama_cleaner.schema import Config as lama_Config |
|
except Exception as e: |
|
lama_cleaner_enable = False |
|
|
|
|
|
from segment_anything import build_sam, SamPredictor, SamAutomaticMaskGenerator |
|
|
|
|
|
import PIL |
|
import requests |
|
import torch |
|
from io import BytesIO |
|
from diffusers import StableDiffusionInpaintPipeline |
|
from huggingface_hub import hf_hub_download |
|
|
|
from util_computer import computer_info |
|
|
|
|
|
from ram_utils import iou, sort_and_deduplicate, relation_classes, MLP, show_anns, ram_show_mask |
|
from ram_train_eval import RamModel, RamPredictor |
|
from mmengine.config import Config as mmengine_Config |
|
|
|
if lama_cleaner_enable: |
|
from lama_cleaner.helper import ( |
|
load_img, |
|
numpy_to_bytes, |
|
resize_max_size, |
|
) |
|
|
|
|
|
import ast |
|
|
|
if kosmos_enable and install_stuff: |
|
os.system("pip install transformers@git+https://github.com/huggingface/transformers.git@main") |
|
|
|
|
|
from kosmos_utils import * |
|
|
|
from util_tencent import getTextTrans |
|
|
|
config_file = 'GroundingDINO/groundingdino/config/GroundingDINO_SwinT_OGC.py' |
|
ckpt_repo_id = "ShilongLiu/GroundingDINO" |
|
ckpt_filenmae = "groundingdino_swint_ogc.pth" |
|
sam_checkpoint = './sam_hq_vit_h.pth' |
|
output_dir = "outputs" |
|
|
|
device = 'cpu' |
|
os.makedirs(output_dir, exist_ok=True) |
|
groundingdino_model = None |
|
sam_device = "cuda" |
|
sam_model = None |
|
|
|
|
|
def get_sam_vit_h_4b8939(): |
|
url = 'https://huggingface.co/Uminosachi/sam-hq/resolve/main/sam_hq_vit_h.pth' |
|
file_path = './sam_hq_vit_h.pth' |
|
|
|
if not os.path.exists(file_path): |
|
logger.info("Downloading sam_vit_h_4b8939.pth...") |
|
response = requests.get(url) |
|
with open(file_path, 'wb') as f: |
|
f.write(response.content) |
|
print('Downloaded sam_vit_h_4b8939.pth') |
|
|
|
get_sam_vit_h_4b8939() |
|
logger.info(f"initialize SAM model...") |
|
sam_device = "cuda" |
|
sam_model = build_sam(checkpoint=sam_checkpoint).to(sam_device) |
|
sam_predictor = SamPredictor(sam_model) |
|
sam_mask_generator = SamAutomaticMaskGenerator(sam_model) |
|
|
|
sam_mask_generator = None |
|
sd_model = None |
|
lama_cleaner_model= None |
|
ram_model = None |
|
kosmos_model = None |
|
kosmos_processor = None |
|
|
|
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) |
|
|
|
|
|
for box, label in zip(boxes, labels): |
|
|
|
box = box * torch.Tensor([W, H, W, H]) |
|
|
|
box[:2] -= box[2:] / 2 |
|
box[2:] += box[:2] |
|
|
|
color = tuple(np.random.randint(0, 255, size=3).tolist()) |
|
|
|
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) |
|
|
|
|
|
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) |
|
|
|
draw.rectangle(bbox, fill=color) |
|
|
|
try: |
|
font = os.path.join(cv2.__path__[0],'qt','fonts','DejaVuSans.ttf') |
|
font_size = 36 |
|
new_font = ImageFont.truetype(font, font_size) |
|
|
|
draw.text((x0+2, y0+2), str(label), font=new_font, fill="white") |
|
except Exception as e: |
|
pass |
|
|
|
mask_draw.rectangle([x0, y0, x1, y1], fill=255, width=6) |
|
|
|
|
|
return image_pil, mask |
|
|
|
def load_image(image_path): |
|
|
|
if isinstance(image_path, PIL.Image.Image): |
|
image_pil = image_path |
|
else: |
|
image_pil = Image.open(image_path).convert("RGB") |
|
|
|
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) |
|
return image_pil, image |
|
|
|
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] |
|
boxes = outputs["pred_boxes"].cpu()[0] |
|
logits.shape[0] |
|
|
|
|
|
logits_filt = logits.clone() |
|
boxes_filt = boxes.clone() |
|
filt_mask = logits_filt.max(dim=1)[0] > box_threshold |
|
logits_filt = logits_filt[filt_mask] |
|
boxes_filt = boxes_filt[filt_mask] |
|
logits_filt.shape[0] |
|
|
|
|
|
tokenlizer = model.tokenizer |
|
tokenized = tokenlizer(caption) |
|
|
|
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)) |
|
|
|
def set_device(): |
|
if os.environ.get('IS_MY_DEBUG') is None: |
|
device = 'cuda' if torch.cuda.is_available() else 'cpu' |
|
else: |
|
device = 'cpu' |
|
print(f'device={device}') |
|
return device |
|
|
|
def load_groundingdino_model(device): |
|
|
|
logger.info(f"initialize groundingdino model...") |
|
groundingdino_model = load_model_hf(config_file, ckpt_repo_id, ckpt_filenmae, device=device) |
|
return groundingdino_model |
|
|
|
|
|
|
|
def load_sam_model(device): |
|
|
|
global sam_model, sam_predictor, sam_mask_generator, sam_device |
|
get_sam_vit_h_4b8939() |
|
logger.info(f"initialize SAM model...") |
|
sam_device = device |
|
sam_model = build_sam(checkpoint=sam_checkpoint).to(sam_device) |
|
sam_predictor = SamPredictor(sam_model) |
|
sam_mask_generator = SamAutomaticMaskGenerator(sam_model) |
|
|
|
def load_sd_model(device): |
|
|
|
global sd_model |
|
logger.info(f"initialize stable-diffusion-inpainting...") |
|
sd_model = None |
|
if os.environ.get('IS_MY_DEBUG') is None: |
|
sd_model = StableDiffusionInpaintPipeline.from_pretrained( |
|
"runwayml/stable-diffusion-inpainting", |
|
revision="fp16", |
|
|
|
torch_dtype=torch.float16, |
|
) |
|
sd_model = sd_model.to(device) |
|
|
|
def load_lama_cleaner_model(device): |
|
|
|
global lama_cleaner_model |
|
logger.info(f"initialize lama_cleaner...") |
|
|
|
lama_cleaner_model = ModelManager( |
|
name='lama', |
|
device=device, |
|
) |
|
|
|
def lama_cleaner_process(image, mask, cleaner_size_limit=1080): |
|
try: |
|
logger.info(f'_______lama_cleaner_process_______1____') |
|
ori_image = image |
|
if mask.shape[0] == image.shape[1] and mask.shape[1] == image.shape[0] and mask.shape[0] != mask.shape[1]: |
|
|
|
logger.info(f'_______lama_cleaner_process_______2____') |
|
ori_image = np.transpose(image[::-1, ...][:, ::-1], axes=(1, 0, 2))[::-1, ...] |
|
logger.info(f'_______lama_cleaner_process_______3____') |
|
image = ori_image |
|
|
|
logger.info(f'_______lama_cleaner_process_______4____') |
|
original_shape = ori_image.shape |
|
logger.info(f'_______lama_cleaner_process_______5____') |
|
interpolation = cv2.INTER_CUBIC |
|
|
|
size_limit = cleaner_size_limit |
|
if size_limit == -1: |
|
logger.info(f'_______lama_cleaner_process_______6____') |
|
size_limit = max(image.shape) |
|
else: |
|
logger.info(f'_______lama_cleaner_process_______7____') |
|
size_limit = int(size_limit) |
|
|
|
logger.info(f'_______lama_cleaner_process_______8____') |
|
config = lama_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, |
|
) |
|
|
|
logger.info(f'_______lama_cleaner_process_______9____') |
|
if config.sd_seed == -1: |
|
config.sd_seed = random.randint(1, 999999999) |
|
|
|
|
|
logger.info(f'_______lama_cleaner_process_______10____') |
|
image = resize_max_size(image, size_limit=size_limit, interpolation=interpolation) |
|
|
|
|
|
|
|
logger.info(f'_______lama_cleaner_process_______11____') |
|
mask = resize_max_size(mask, size_limit=size_limit, interpolation=interpolation) |
|
|
|
|
|
logger.info(f'_______lama_cleaner_process_______12____') |
|
res_np_img = lama_cleaner_model(image, mask, config) |
|
logger.info(f'_______lama_cleaner_process_______13____') |
|
torch.cuda.empty_cache() |
|
|
|
logger.info(f'_______lama_cleaner_process_______14____') |
|
image = Image.open(io.BytesIO(numpy_to_bytes(res_np_img, 'png'))) |
|
logger.info(f'_______lama_cleaner_process_______15____') |
|
except Exception as e: |
|
logger.info(f'lama_cleaner_process[Error]:' + str(e)) |
|
image = None |
|
return image |
|
|
|
class Ram_Predictor(RamPredictor): |
|
def __init__(self, config, device='cpu'): |
|
self.config = config |
|
self.device = torch.device(device) |
|
self._build_model() |
|
|
|
def _build_model(self): |
|
self.model = RamModel(**self.config.model).to(self.device) |
|
if self.config.load_from is not None: |
|
self.model.load_state_dict(torch.load(self.config.load_from, map_location=self.device)) |
|
self.model.train() |
|
|
|
def load_ram_model(device): |
|
|
|
global ram_model |
|
if os.environ.get('IS_MY_DEBUG') is not None: |
|
return |
|
model_path = "./checkpoints/ram_epoch12.pth" |
|
ram_config = dict( |
|
model=dict( |
|
pretrained_model_name_or_path='bert-base-uncased', |
|
load_pretrained_weights=False, |
|
num_transformer_layer=2, |
|
input_feature_size=256, |
|
output_feature_size=768, |
|
cls_feature_size=512, |
|
num_relation_classes=56, |
|
pred_type='attention', |
|
loss_type='multi_label_ce', |
|
), |
|
load_from=model_path, |
|
) |
|
ram_config = mmengine_Config(ram_config) |
|
ram_model = Ram_Predictor(ram_config, device) |
|
|
|
|
|
def draw_selected_mask(mask, draw): |
|
color = (255, 0, 0, 153) |
|
nonzero_coords = np.transpose(np.nonzero(mask)) |
|
for coord in nonzero_coords: |
|
draw.point(coord[::-1], fill=color) |
|
|
|
def draw_object_mask(mask, draw): |
|
color = (0, 0, 255, 153) |
|
nonzero_coords = np.transpose(np.nonzero(mask)) |
|
for coord in nonzero_coords: |
|
draw.point(coord[::-1], fill=color) |
|
|
|
def create_title_image(word1, word2, word3, width, font_path='./assets/OpenSans-Bold.ttf'): |
|
|
|
color_red = (255, 0, 0) |
|
color_black = (0, 0, 0) |
|
color_blue = (0, 0, 255) |
|
|
|
|
|
font_size = 40 |
|
|
|
|
|
image = Image.new('RGB', (width, 60), (255, 255, 255)) |
|
|
|
try: |
|
|
|
font = ImageFont.truetype(font_path, font_size) |
|
|
|
|
|
while True: |
|
|
|
draw = ImageDraw.Draw(image) |
|
|
|
word_spacing = font_size / 2 |
|
|
|
x_offset = word_spacing |
|
draw.text((x_offset, 0), word1, color_red, font=font) |
|
x_offset += font.getsize(word1)[0] + word_spacing |
|
draw.text((x_offset, 0), word2, color_black, font=font) |
|
x_offset += font.getsize(word2)[0] + word_spacing |
|
draw.text((x_offset, 0), word3, color_blue, font=font) |
|
|
|
word_sizes = [font.getsize(word) for word in [word1, word2, word3]] |
|
total_width = sum([size[0] for size in word_sizes]) + word_spacing * 3 |
|
|
|
|
|
if total_width <= width: |
|
break |
|
|
|
|
|
font_size -= 1 |
|
image = Image.new('RGB', (width, 50), (255, 255, 255)) |
|
font = ImageFont.truetype(font_path, font_size) |
|
draw = None |
|
except Exception as e: |
|
pass |
|
|
|
return image |
|
|
|
def concatenate_images_vertical(image1, image2): |
|
|
|
width1, height1 = image1.size |
|
width2, height2 = image2.size |
|
|
|
|
|
new_image = Image.new('RGBA', (max(width1, width2), height1 + height2)) |
|
|
|
|
|
new_image.paste(image1, (0, 0)) |
|
|
|
|
|
new_image.paste(image2, (0, height1)) |
|
|
|
return new_image |
|
|
|
def relate_anything(input_image, k): |
|
logger.info(f'relate_anything_1_{input_image.size}_') |
|
w, h = input_image.size |
|
max_edge = 1500 |
|
if w > max_edge or h > max_edge: |
|
ratio = max(w, h) / max_edge |
|
new_size = (int(w / ratio), int(h / ratio)) |
|
input_image.thumbnail(new_size) |
|
|
|
logger.info(f'relate_anything_2_') |
|
|
|
pil_image = input_image.convert('RGBA') |
|
image = np.array(input_image) |
|
sam_masks = sam_mask_generator.generate(image) |
|
filtered_masks = sort_and_deduplicate(sam_masks) |
|
|
|
logger.info(f'relate_anything_3_') |
|
feat_list = [] |
|
for fm in filtered_masks: |
|
feat = torch.Tensor(fm['feat']).unsqueeze(0).unsqueeze(0).to(device) |
|
feat_list.append(feat) |
|
feat = torch.cat(feat_list, dim=1).to(device) |
|
matrix_output, rel_triplets = ram_model.predict(feat) |
|
|
|
logger.info(f'relate_anything_4_') |
|
pil_image_list = [] |
|
for i, rel in enumerate(rel_triplets[:k]): |
|
s,o,r = int(rel[0]),int(rel[1]),int(rel[2]) |
|
relation = relation_classes[r] |
|
|
|
mask_image = Image.new('RGBA', pil_image.size, color=(0, 0, 0, 0)) |
|
mask_draw = ImageDraw.Draw(mask_image) |
|
|
|
draw_selected_mask(filtered_masks[s]['segmentation'], mask_draw) |
|
draw_object_mask(filtered_masks[o]['segmentation'], mask_draw) |
|
|
|
current_pil_image = pil_image.copy() |
|
current_pil_image.alpha_composite(mask_image) |
|
|
|
title_image = create_title_image('Red', relation, 'Blue', current_pil_image.size[0]) |
|
concate_pil_image = concatenate_images_vertical(current_pil_image, title_image) |
|
pil_image_list.append(concate_pil_image) |
|
|
|
logger.info(f'relate_anything_5_{len(pil_image_list)}') |
|
return pil_image_list |
|
|
|
mask_source_draw = "draw a mask on input image" |
|
mask_source_segment = "type what to detect below" |
|
|
|
def get_time_cost(run_task_time, time_cost_str): |
|
now_time = int(time.time()*1000) |
|
if run_task_time == 0: |
|
time_cost_str = 'start' |
|
else: |
|
if time_cost_str != '': |
|
time_cost_str += f'-->' |
|
time_cost_str += f'{now_time - run_task_time}' |
|
run_task_time = now_time |
|
return run_task_time, time_cost_str |
|
|
|
def run_anything_task(input_image, text_prompt, task_type, inpaint_prompt, box_threshold, text_threshold, |
|
iou_threshold, inpaint_mode, mask_source_radio, remove_mode, remove_mask_extend, num_relation, kosmos_input, cleaner_size_limit=1080): |
|
|
|
text_prompt = getTextTrans(text_prompt, source='zh', target='en') |
|
inpaint_prompt = getTextTrans(inpaint_prompt, source='zh', target='en') |
|
|
|
run_task_time = 0 |
|
time_cost_str = '' |
|
run_task_time, time_cost_str = get_time_cost(run_task_time, time_cost_str) |
|
|
|
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!😂😂😂😂'), time_cost_str, gr.Textbox.update(visible=(time_cost_str !='')), None, None, None |
|
|
|
if input_image is None: |
|
return [], gr.Gallery.update(label='Please upload a image!😂😂😂😂'), time_cost_str, gr.Textbox.update(visible=(time_cost_str !='')), None, None, None |
|
|
|
file_temp = int(time.time()) |
|
logger.info(f'run_anything_task_002/{device}_[{file_temp}]_{task_type}/{inpaint_mode}/[{mask_source_radio}]/{remove_mode}/{remove_mask_extend}_[{text_prompt}]/[{inpaint_prompt}]___1_') |
|
|
|
output_images = [] |
|
|
|
image_pil, image = load_image(input_image.convert("RGB")) |
|
input_img = input_image |
|
run_task_time, time_cost_str = get_time_cost(run_task_time, time_cost_str) |
|
|
|
size = image_pil.size |
|
H, W = size[1], size[0] |
|
|
|
|
|
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!") |
|
|
|
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_anything_task_[{file_temp}]_{task_type}_[{text_prompt}]_1___{groundingdino_device}/[No objects detected, please try others.]_') |
|
return [], gr.Gallery.update(label='No objects detected, please try others.😂😂😂😂'), time_cost_str, gr.Textbox.update(visible=(time_cost_str !='')), None, None, None |
|
boxes_filt_ori = copy.deepcopy(boxes_filt) |
|
|
|
logger.info(f'run_anything_task_[{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_img) |
|
if sam_predictor: |
|
sam_predictor.set_image(image) |
|
|
|
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] |
|
|
|
if sam_predictor: |
|
boxes_filt = boxes_filt.to(sam_device) |
|
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, |
|
) |
|
|
|
assert sam_checkpoint, 'sam_checkpoint is not found!' |
|
else: |
|
masks = torch.zeros(len(boxes_filt), 1, H, W) |
|
mask_count = 0 |
|
for box in boxes_filt: |
|
masks[mask_count, 0, int(box[1]):int(box[3]), int(box[0]):int(box[2])] = 1 |
|
mask_count += 1 |
|
masks = torch.where(masks > 0, True, False) |
|
run_mode = "rectangle" |
|
|
|
|
|
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.cpu().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") |
|
plt.clf() |
|
plt.close('all') |
|
segment_image_result = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB) |
|
os.remove(image_path) |
|
output_images.append(Image.fromarray(segment_image_result)) |
|
run_task_time, time_cost_str = get_time_cost(run_task_time, time_cost_str) |
|
|
|
print(sam_predictor) |
|
|
|
if inpaint_prompt.strip() == '' and mask_source_radio == mask_source_segment: |
|
task_type = 'remove' |
|
|
|
logger.info(f'run_anything_task_[{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) |
|
output_images.append(mask_pil.convert("RGB")) |
|
return mask_pil |
|
|
|
def change_radio_display(task_type, mask_source_radio): |
|
text_prompt_visible = True |
|
inpaint_prompt_visible = False |
|
mask_source_radio_visible = False |
|
num_relation_visible = False |
|
|
|
image_gallery_visible = True |
|
kosmos_input_visible = False |
|
kosmos_output_visible = False |
|
kosmos_text_output_visible = False |
|
|
|
if task_type == "Kosmos-2": |
|
if kosmos_enable: |
|
text_prompt_visible = False |
|
image_gallery_visible = False |
|
kosmos_input_visible = True |
|
kosmos_output_visible = True |
|
kosmos_text_output_visible = True |
|
|
|
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 |
|
if task_type == "relate anything": |
|
text_prompt_visible = False |
|
num_relation_visible = True |
|
|
|
return (gr.Textbox.update(visible=text_prompt_visible), |
|
gr.Textbox.update(visible=inpaint_prompt_visible), |
|
gr.Radio.update(visible=mask_source_radio_visible), |
|
gr.Slider.update(visible=num_relation_visible), |
|
gr.Gallery.update(visible=image_gallery_visible), |
|
gr.Radio.update(visible=kosmos_input_visible), |
|
gr.Image.update(visible=kosmos_output_visible), |
|
gr.HighlightedText.update(visible=kosmos_text_output_visible)) |
|
|
|
def get_model_device(module): |
|
try: |
|
if module is None: |
|
return 'None' |
|
if isinstance(module, torch.nn.DataParallel): |
|
module = module.module |
|
for submodule in module.children(): |
|
if hasattr(submodule, "_parameters"): |
|
parameters = submodule._parameters |
|
if "weight" in parameters: |
|
return parameters["weight"].device |
|
return 'UnKnown' |
|
except Exception as e: |
|
return 'Error' |
|
|
|
def main_gradio(args): |
|
block = gr.Blocks().queue() |
|
with block: |
|
with gr.Row(): |
|
with gr.Column(): |
|
task_types = ["detection"] |
|
if sam_enable: |
|
task_types.append("segment") |
|
if inpainting_enable: |
|
task_types.append("inpainting") |
|
if lama_cleaner_enable: |
|
task_types.append("remove") |
|
if ram_enable: |
|
task_types.append("relate anything") |
|
if kosmos_enable: |
|
task_types.append("Kosmos-2") |
|
|
|
input_image = gr.Image(source='upload', elem_id="image_upload", tool='sketch', type='pil', label="Upload", |
|
height=512, brush_color='#00FFFF', mask_opacity=0.6) |
|
task_type = gr.Radio(task_types, value="detection", |
|
label='Task type', visible=True) |
|
mask_source_radio = gr.Radio([mask_source_draw, mask_source_segment], |
|
value=mask_source_segment, label="Mask from", |
|
visible=False) |
|
text_prompt = gr.Textbox(label="Detection Prompt[To detect multiple objects, seperating each 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) |
|
num_relation = gr.Slider(label="How many relations do you want to see", minimum=1, maximum=20, value=5, step=1, visible=False) |
|
|
|
kosmos_input = gr.Radio(["Brief", "Detailed"], label="Kosmos Description Type", value="Brief", visible=False) |
|
|
|
run_button = gr.Button(label="Run", visible=True) |
|
with gr.Accordion("Advanced options", open=False) as advanced_options: |
|
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.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(): |
|
image_gallery = gr.Gallery(label="result images", show_label=True, elem_id="gallery", height=512, visible=True |
|
).style(preview=True, columns=[5], object_fit="scale-down", height="auto") |
|
time_cost = gr.Textbox(label="Time cost by step (ms):", visible=False, interactive=False) |
|
|
|
kosmos_output = gr.Image(type="pil", label="result images", visible=False) |
|
kosmos_text_output = gr.HighlightedText( |
|
label="Generated Description", |
|
combine_adjacent=False, |
|
show_legend=True, |
|
visible=False, |
|
).style(color_map=color_map) |
|
|
|
selected = gr.Number(-1, show_label=False, placeholder="Selected", visible=False) |
|
|
|
|
|
entity_output = gr.Textbox(visible=False) |
|
|
|
|
|
def get_text_span_label(evt: gr.SelectData): |
|
if evt.value[-1] is None: |
|
return -1 |
|
return int(evt.value[-1]) |
|
|
|
kosmos_text_output.select(get_text_span_label, None, selected) |
|
|
|
|
|
def update_output_image(img_input, image_output, entities, idx): |
|
entities = ast.literal_eval(entities) |
|
updated_image = draw_entity_boxes_on_image(img_input, entities, entity_index=idx) |
|
return updated_image |
|
selected.change(update_output_image, [kosmos_output, kosmos_output, entity_output, selected], [kosmos_output]) |
|
|
|
run_button.click(fn=run_anything_task, 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, num_relation, kosmos_input], |
|
outputs=[image_gallery, image_gallery, time_cost, time_cost, kosmos_output, kosmos_text_output, entity_output], show_progress=True, queue=True) |
|
|
|
mask_source_radio.change(fn=change_radio_display, inputs=[task_type, mask_source_radio], |
|
outputs=[text_prompt, inpaint_prompt, mask_source_radio, num_relation]) |
|
task_type.change(fn=change_radio_display, inputs=[task_type, mask_source_radio], |
|
outputs=[text_prompt, inpaint_prompt, mask_source_radio, num_relation, |
|
image_gallery, kosmos_input, kosmos_output, kosmos_text_output |
|
]) |
|
|
|
DESCRIPTION = f'### This demo from [Grounded-Segment-Anything](https://github.com/IDEA-Research/Grounded-Segment-Anything). <br>' |
|
if lama_cleaner_enable: |
|
DESCRIPTION += f'Remove(cleaner) from [lama-cleaner](https://github.com/Sanster/lama-cleaner). <br>' |
|
if kosmos_enable: |
|
DESCRIPTION += f'Kosmos-2 from [Kosmos-2](https://github.com/microsoft/unilm/tree/master/kosmos-2). <br>' |
|
if ram_enable: |
|
DESCRIPTION += f'RAM from [RelateAnything](https://github.com/Luodian/RelateAnything). <br>' |
|
DESCRIPTION += f'Thanks for their excellent work.' |
|
DESCRIPTION += f'<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings. \ |
|
<a href="https://huggingface.co/spaces/yizhangliu/Grounded-Segment-Anything?duplicate=true"><img style="display: inline; margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space" /></a></p>' |
|
gr.Markdown(DESCRIPTION) |
|
|
|
print(f'device = {device}') |
|
print(f'torch.cuda.is_available = {torch.cuda.is_available()}') |
|
computer_info() |
|
block.launch(server_name='0.0.0.0', server_port=args.port, debug=args.debug, share=args.share) |
|
|
|
import signal |
|
import json |
|
from datetime import date, datetime, timedelta |
|
from gevent import pywsgi |
|
import base64 |
|
|
|
def imgFile_to_base64(image_file): |
|
with open(image_file, "rb") as f: |
|
im_bytes = f.read() |
|
im_b64_encode = base64.b64encode(im_bytes) |
|
im_b64 = im_b64_encode.decode("utf8") |
|
return im_b64 |
|
|
|
def base64_to_bytes(im_b64): |
|
im_b64_encode = im_b64.encode("utf-8") |
|
im_bytes = base64.b64decode(im_b64_encode) |
|
return im_bytes |
|
|
|
def base64_to_PILImage(im_b64): |
|
im_bytes = base64_to_bytes(im_b64) |
|
pil_img = Image.open(io.BytesIO(im_bytes)) |
|
return pil_img |
|
|
|
class API_Starter: |
|
def __init__(self): |
|
from flask import Flask, request, jsonify, make_response |
|
from flask_cors import CORS, cross_origin |
|
import logging |
|
|
|
app = Flask(__name__) |
|
app.logger.setLevel(logging.ERROR) |
|
CORS(app, supports_credentials=True, resources={r"/*": {"origins": "*"}}) |
|
|
|
@app.route('/imgCLeaner', methods=['GET', 'POST']) |
|
@cross_origin() |
|
def processAssist(): |
|
if request.method == 'GET': |
|
ret_json = {'code': -1, 'reason':'no support to get'} |
|
elif request.method == 'POST': |
|
request_data = request.data.decode('utf-8') |
|
data = json.loads(request_data) |
|
result = self.handle_data(data) |
|
if result is None: |
|
ret_json = {'code': -2, 'reason':'handle error'} |
|
else: |
|
ret_json = {'code': 0, 'result':result} |
|
return jsonify(ret_json) |
|
|
|
self.app = app |
|
now_time = datetime.now().strftime('%Y%m%d_%H%M%S') |
|
logger.add(f'./logs/logger_[{args.port}]_{now_time}.log') |
|
signal.signal(signal.SIGINT, self.signal_handler) |
|
|
|
def handle_data(self, data): |
|
im_b64 = data['img'] |
|
img = base64_to_PILImage(im_b64) |
|
remove_texts = data['remove_texts'] |
|
remove_mask_extend = data['mask_extend'] |
|
results = run_anything_task(input_image = img, |
|
text_prompt = f"{remove_texts}", |
|
task_type = 'remove', |
|
inpaint_prompt = '', |
|
box_threshold = 0.3, |
|
text_threshold = 0.25, |
|
iou_threshold = 0.8, |
|
inpaint_mode = "merge", |
|
mask_source_radio = "type what to detect below", |
|
remove_mode = "rectangle", |
|
remove_mask_extend = f"{remove_mask_extend}", |
|
num_relation = 5, |
|
kosmos_input = None, |
|
cleaner_size_limit = -1, |
|
) |
|
output_images = results[0] |
|
if output_images is None: |
|
return None |
|
ret_json_images = [] |
|
file_temp = int(time.time()) |
|
count = 0 |
|
output_images = output_images[-1:] |
|
for image_pil in output_images: |
|
try: |
|
img_format = image_pil.format.lower() |
|
except Exception as e: |
|
img_format = 'png' |
|
image_path = os.path.join(output_dir, f"api_images_{file_temp}_{count}.{img_format}") |
|
count += 1 |
|
try: |
|
image_pil.save(image_path) |
|
except Exception as e: |
|
Image.fromarray(image_pil).save(image_path) |
|
im_b64 = imgFile_to_base64(image_path) |
|
ret_json_images.append(im_b64) |
|
os.remove(image_path) |
|
data = { |
|
'imgs': ret_json_images, |
|
} |
|
return data |
|
|
|
def signal_handler(self, signal, frame): |
|
print('\nSignal Catched! You have just type Ctrl+C!') |
|
sys.exit(0) |
|
|
|
def run(self): |
|
from gevent import pywsgi |
|
logger.info(f'\nargs={args}\n') |
|
computer_info() |
|
print(f"Start a api server: http://0.0.0.0:{args.port}/imgCLeaner") |
|
server = pywsgi.WSGIServer(('0.0.0.0', args.port), self.app) |
|
server.serve_forever() |
|
|
|
device = set_device() |
|
|
|
groundingdino_model = load_groundingdino_model('cuda:0') |
|
load_sam_model("cuda:0") |
|
|
|
load_sd_model("cuda:0") |
|
|
|
load_lama_cleaner_model("cuda:0") |
|
|
|
|
|
|
|
|
|
def expand_white_pixels(input_pil, expand_by=1): |
|
|
|
grayscale = input_pil.convert('L') |
|
|
|
|
|
binary_mask = np.array(grayscale) > 245 |
|
|
|
|
|
dilated_mask = binary_dilation(binary_mask, iterations=expand_by) |
|
|
|
|
|
expanded_image = Image.fromarray(np.uint8(dilated_mask * 255)) |
|
|
|
return expanded_image |
|
|
|
def just_fucking_get_sd_mask(input_pil, prompt, expand_by=10): |
|
raw_mask = run_anything_task(input_pil, prompt, "inpainting", "", 0.3, 0.25, 0.8, "merge", "type what to detect below", "segment", "10", 5, "Brief") |
|
expanded_mask = expand_white_pixels(raw_mask, expand_by=expand_by) |
|
|
|
return expanded_mask |
|
|
|
S3_REGION = "fra1" |
|
S3_ACCESS_ID = "0RN7BZXS59HYSBD3VB79" |
|
S3_ACCESS_SECRET = "hfSPgBlWl5jsGHa2xuByVkSpancgVeA2CVQf2EMp" |
|
S3_ENDPOINT_URL = "https://s3.solarcom.ch" |
|
S3_BUCKET_NAME = "pissnelke" |
|
|
|
import boto3 |
|
|
|
s3_session = boto3.session.Session() |
|
s3 = s3_session.client( |
|
service_name="s3", |
|
region_name=S3_REGION, |
|
aws_access_key_id=S3_ACCESS_ID, |
|
aws_secret_access_key=S3_ACCESS_SECRET, |
|
endpoint_url=S3_ENDPOINT_URL, |
|
) |
|
|
|
class EndpointHandler(): |
|
def __init__(self, path=""): |
|
|
|
os.environ['path'] = path |
|
print("running apt-get update && apt-get install ffmpeg libsm6 libxext6 -y") |
|
command = "apt-get update && apt-get install ffmpeg libsm6 libxext6 -y" |
|
process = subprocess.Popen( |
|
command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE) |
|
stdout, stderr = process.communicate() |
|
print("ran apt-get update && apt-get install ffmpeg libsm6 libxext6 -y") |
|
print("path", path) |
|
|
|
def __call__(self, data): |
|
original_image_res = requests.get(data.get("original_link")) |
|
original_pil = Image.open(BytesIO(original_image_res.content)) |
|
|
|
with_small_tits = data.get("with_small_tits", False) |
|
|
|
with_big_tits = data.get("with_big_tits", False) |
|
|
|
nude_pils = get_nude(get_mask_function=just_fucking_get_sd_mask, cfg_scale=data.get("cfg_scale"), generate_max_size=data.get("generate_max_size"), original_max_size=data.get( |
|
"original_max_size"), original_pil=original_pil, positive_prompt=data.get("positive_prompt"), steps=data.get("steps"), with_small_tits=with_small_tits, with_big_tits=with_big_tits) |
|
|
|
filenames = [] |
|
|
|
for image in nude_pils: |
|
byte_arr = io.BytesIO() |
|
image.save(byte_arr, format='PNG') |
|
byte_arr = byte_arr.getvalue() |
|
|
|
random_string = ''.join(random.choice( |
|
string.ascii_letters + string.digits) for i in range(20)) |
|
image_filename = random_string + ".jpeg" |
|
|
|
s3.put_object(Body=byte_arr, Bucket=S3_BUCKET_NAME, |
|
Key=image_filename) |
|
|
|
filenames.append(image_filename) |
|
|
|
return { |
|
"filenames": filenames |
|
} |
|
|
|
print(EndpointHandler()({ |
|
"original_link": "https://www.shutterstock.com/image-photo/attractive-confident-young-woman-posing-600nw-2185228917.jpg" |
|
})) |
|
|