159 lines
4.9 KiB
Python
159 lines
4.9 KiB
Python
# Based on https://github.com/xuebinqin/DIS/blob/main/Colab_Demo.ipynb
|
|
from PIL import Image
|
|
import numpy as np
|
|
import torch
|
|
from torch.autograd import Variable
|
|
from torchvision import transforms
|
|
import torch.nn.functional as F
|
|
|
|
device = None
|
|
ISNetDIS = None
|
|
normalize = None
|
|
im_preprocess = None
|
|
hypar = None
|
|
net = None
|
|
|
|
|
|
def init(saved_models_path):
|
|
global device, ISNetDIS, normalize, im_preprocess, hypar, net
|
|
|
|
print("### ComfyUI-Background-Replacement: Initializing segmenter...")
|
|
|
|
from .models.isnet import ISNetDIS
|
|
from .data_loader_cache import normalize, im_preprocess
|
|
|
|
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
|
# ISNetDIS = models.ISNetDIS
|
|
# normalize = data_loader_cache.normalize
|
|
# im_preprocess = data_loader_cache.im_preprocess
|
|
|
|
# Set Parameters
|
|
hypar = {} # paramters for inferencing
|
|
|
|
# load trained weights from this path
|
|
hypar["model_path"] = saved_models_path # "./saved_models"
|
|
# name of the to-be-loaded weights
|
|
hypar["restore_model"] = "isnet-general-use.pth"
|
|
# indicate if activate intermediate feature supervision
|
|
hypar["interm_sup"] = False
|
|
|
|
# choose floating point accuracy --
|
|
# indicates "half" or "full" accuracy of float number
|
|
hypar["model_digit"] = "full"
|
|
hypar["seed"] = 0
|
|
|
|
# cached input spatial resolution, can be configured into different size
|
|
hypar["cache_size"] = [1024, 1024]
|
|
|
|
# data augmentation parameters ---
|
|
# mdoel input spatial size, usually use the same value hypar["cache_size"], which means we don't further resize the images
|
|
hypar["input_size"] = [1024, 1024]
|
|
# random crop size from the input, it is usually set as smaller than hypar["cache_size"], e.g., [920,920] for data augmentation
|
|
hypar["crop_size"] = [1024, 1024]
|
|
|
|
hypar["model"] = ISNetDIS()
|
|
|
|
# Build Model
|
|
net = build_model(hypar, device)
|
|
|
|
|
|
class GOSNormalize(object):
|
|
'''
|
|
Normalize the Image using torch.transforms
|
|
'''
|
|
|
|
def __init__(self, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]):
|
|
self.mean = mean
|
|
self.std = std
|
|
|
|
def __call__(self, image):
|
|
image = normalize(image, self.mean, self.std)
|
|
return image
|
|
|
|
|
|
transform = transforms.Compose(
|
|
[GOSNormalize([0.5, 0.5, 0.5], [1.0, 1.0, 1.0])])
|
|
|
|
|
|
def load_image(im_pil, hypar):
|
|
# Convert PIL Image to NumPy array
|
|
im = np.array(im_pil)
|
|
# Preprocess the image using im_preprocess function
|
|
im, im_shp = im_preprocess(im, hypar["cache_size"])
|
|
# Normalize pixel values to the range [0, 1]
|
|
im = torch.divide(im, 255.0)
|
|
# Convert image shape to a torch tensor
|
|
shape = torch.from_numpy(np.array(im_shp))
|
|
# Make a batch of image and shape, then apply the specified transformations
|
|
return transform(im).unsqueeze(0), shape.unsqueeze(0)
|
|
|
|
|
|
def build_model(hypar, device):
|
|
net = hypar["model"] # GOSNETINC(3,1)
|
|
|
|
# convert to half precision
|
|
if (hypar["model_digit"] == "half"):
|
|
net.half()
|
|
for layer in net.modules():
|
|
if isinstance(layer, nn.BatchNorm2d):
|
|
layer.float()
|
|
|
|
net.to(device)
|
|
|
|
if (hypar["restore_model"] != ""):
|
|
net.load_state_dict(torch.load(
|
|
hypar["model_path"]+"/"+hypar["restore_model"], map_location=device))
|
|
net.to(device)
|
|
net.eval()
|
|
return net
|
|
|
|
|
|
def predict(net, inputs_val, shapes_val, hypar, device):
|
|
'''
|
|
Given an Image, predict the mask
|
|
'''
|
|
net.eval()
|
|
|
|
if (hypar["model_digit"] == "full"):
|
|
inputs_val = inputs_val.type(torch.FloatTensor)
|
|
else:
|
|
inputs_val = inputs_val.type(torch.HalfTensor)
|
|
|
|
inputs_val_v = Variable(inputs_val, requires_grad=False).to(
|
|
device) # wrap inputs in Variable
|
|
|
|
ds_val = net(inputs_val_v)[0] # list of 6 results
|
|
|
|
# B x 1 x H x W # we want the first one which is the most accurate prediction
|
|
pred_val = ds_val[0][0, :, :, :]
|
|
|
|
# recover the prediction spatial size to the orignal image size
|
|
pred_val = torch.squeeze(F.upsample(torch.unsqueeze(
|
|
pred_val, 0), (shapes_val[0][0], shapes_val[0][1]), mode='bilinear'))
|
|
|
|
ma = torch.max(pred_val)
|
|
mi = torch.min(pred_val)
|
|
pred_val = (pred_val-mi)/(ma-mi) # max = 1
|
|
|
|
if device == 'cuda':
|
|
torch.cuda.empty_cache()
|
|
# it is the mask we need
|
|
return (pred_val.detach().cpu().numpy()*255).astype(np.uint8)
|
|
|
|
|
|
def segment(image):
|
|
# Load the image and get the image tensor and original size
|
|
image_tensor, orig_size = load_image(image, hypar)
|
|
# Predict the segmentation mask using the neural network
|
|
mask = predict(net, image_tensor, orig_size, hypar, device)
|
|
# Convert the mask to a PIL Image and grayscale
|
|
mask = Image.fromarray(mask).convert('L')
|
|
# Convert the original image to RGB mode
|
|
im_rgb = image.convert("RGB")
|
|
# Create a copy of the RGB image
|
|
cropped = im_rgb.copy()
|
|
# Apply the alpha channel (transparency) using the predicted mask
|
|
cropped.putalpha(mask)
|
|
# Return the cropped image and the mask
|
|
return [cropped, mask]
|