linxy97/CustomCodeForRMBG

1import torch, os2import torch.nn.functional as F3from torchvision.transforms.functional import normalize4import numpy as np5from transformers import Pipeline6from skimage import io7from PIL import Image8 9 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, orig_im: Image, model_input_size: list = [1024, 1024]):28        # preprocess the input29        orig_im_size = orig_im.shape[0:2]30        image = self.preprocess_image(orig_im, model_input_size).to(self.device)31        inputs = {32            "orig_im": orig_im,33            "image": image,34            "orig_im_size": orig_im_size,35        }36        return inputs37 38    def _forward(self, inputs):39        result = self.model(inputs.pop("image"))40        inputs["result"] = result41        return inputs42 43    def postprocess(self, inputs, out_name=""):44        result = inputs.pop("result")45        orig_im_size = inputs.pop("orig_im_size")46        orig_image = inputs.pop("orig_im")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            no_bg_image.paste(orig_image, mask=pil_im)53            no_bg_image.save(out_name)54        else:55            return result_image56 57    # utilities functions58    def preprocess_image(59        self, im: np.ndarray, model_input_size: list = [1024, 1024]60    ) -> torch.Tensor:61        # same as utilities.py with minor modification62        if len(im.shape) < 3:63            im = im[:, :, np.newaxis]64        # orig_im_size=im.shape[0:2]65        im_tensor = torch.tensor(im, dtype=torch.float32).permute(2, 0, 1)66        im_tensor = F.interpolate(67            torch.unsqueeze(im_tensor, 0), size=model_input_size, mode="bilinear"68        ).type(torch.uint8)69        image = torch.divide(im_tensor, 255.0)70        image = normalize(image, [0.5, 0.5, 0.5], [1.0, 1.0, 1.0])71        return image72 73    def postprocess_image(self, result: torch.Tensor, im_size: list) -> np.ndarray:74        result = torch.squeeze(F.interpolate(result, size=im_size, mode="bilinear"), 0)75        ma = torch.max(result)76        mi = torch.min(result)77        result = (result - mi) / (ma - mi)78        im_array = (result * 255).permute(1, 2, 0).cpu().data.numpy().astype(np.uint8)79        im_array = np.squeeze(im_array)80        return im_array81