not-lain/CustomCodeForRMBG

1 2import torch, os3import torch.nn.functional as F4from torchvision.transforms.functional import normalize5import numpy as np6from transformers import Pipeline7from skimage import io8from PIL import Image9 10class RMBGPipe(Pipeline):11  def __init__(self,**kwargs):12    Pipeline.__init__(self,**kwargs)13    self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")14    self.model.to(self.device)15    self.model.eval()16 17  def _sanitize_parameters(self, **kwargs):18    # parse parameters19    preprocess_kwargs = {}20    postprocess_kwargs = {}21    if "model_input_size" in kwargs : 22      preprocess_kwargs["model_input_size"] = kwargs["model_input_size"]23    if "out_name" in kwargs: 24      postprocess_kwargs["out_name"] = kwargs["out_name"]25    return preprocess_kwargs, {}, postprocess_kwargs26 27  def preprocess(self,im_path:str,model_input_size: list=[1024,1024]):28      # preprocess the input 29      orig_im = io.imread(im_path)30      orig_im_size = orig_im.shape[0:2]31      image = self.preprocess_image(orig_im, model_input_size).to(self.device)32      inputs = {33          "image":image,34          "orig_im_size":orig_im_size,35          "im_path" : im_path36      }37      return inputs38 39  def _forward(self,inputs):40    result = self.model(inputs.pop("image"))41    inputs["result"] = result42    return inputs43  def postprocess(self,inputs,out_name = ""):44    result = inputs.pop("result")45    orig_im_size = inputs.pop("orig_im_size")46    im_path = inputs.pop("im_path")47    result_image = self.postprocess_image(result[0][0], orig_im_size)48    if out_name != "" : 49      # if out_name is specified we save the image using that name50      pil_im = Image.fromarray(result_image)51      no_bg_image = Image.new("RGBA", pil_im.size, (0,0,0,0))52      orig_image = Image.open(im_path)53      no_bg_image.paste(orig_image, mask=pil_im)54      no_bg_image.save(out_name)55    else : 56      return result_image57 58  # utilities functions59  def preprocess_image(self,im: np.ndarray, model_input_size: list=[1024,1024]) -> torch.Tensor:60    # same as utilities.py with minor modification61    if len(im.shape) < 3:62        im = im[:, :, np.newaxis]63    # orig_im_size=im.shape[0:2]64    im_tensor = torch.tensor(im, dtype=torch.float32).permute(2,0,1)65    im_tensor = F.interpolate(torch.unsqueeze(im_tensor,0), size=model_input_size, mode='bilinear').type(torch.uint8)66    image = torch.divide(im_tensor,255.0)67    image = normalize(image,[0.5,0.5,0.5],[1.0,1.0,1.0])68    return image69  def postprocess_image(self,result: torch.Tensor, im_size: list)-> np.ndarray:70      result = torch.squeeze(F.interpolate(result, size=im_size, mode='bilinear') ,0)71      ma = torch.max(result)72      mi = torch.min(result)73      result = (result-mi)/(ma-mi)74      im_array = (result*255).permute(1,2,0).cpu().data.numpy().astype(np.uint8)75      im_array = np.squeeze(im_array)76      return im_array77