import torch import torch.nn.functional as F import torchvision.transforms as T from bert.modeling_bert import BertModel from lib import segmentation import numpy as np class FaceSegment: def __init__( self, model_path: str, embedding_path: str = None, device: str = "cpu", img_size: int = 480, ): self.device = torch.device(device) self.transform = T.Compose([ T.Resize(img_size), T.ToTensor(), T.Normalize( mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] ) ]) token_ids = [101, 2529, 2132, 102] max_len = 20 padded = token_ids + [0] * (max_len - len(token_ids)) attn_mask = [1] * len(token_ids) + [0] * (max_len - len(token_ids)) self.sent_tensor = torch.tensor([padded], dtype=torch.long, device=self.device) self.mask_tensor = torch.tensor([attn_mask], dtype=torch.long, device=self.device) self.l_mask4seg = self.mask_tensor.unsqueeze(-1) class Opts: swin_type = "base" window12 = True mha = "" fusion_drop= 0.0 self.model_seg = segmentation.lavt(pretrained="", args=Opts()) self.model_seg.load_state_dict(torch.load(model_path, map_location="cpu")["model"]) self.model_seg.to(self.device).eval() if embedding_path: emb_data = torch.load(embedding_path, map_location=self.device) self.embedding = emb_data["embedding"].to(self.device) else: bert_ckpt = torch.load(model_path, map_location="cpu")["bert_model"] self.bert = BertModel.from_pretrained("bert-base-uncased") self.bert.pooler = None self.bert.load_state_dict(bert_ckpt) self.bert.to(self.device).eval() with torch.no_grad(): last_hidden = self.bert(self.sent_tensor, attention_mask=self.mask_tensor)[0] self.embedding = last_hidden.permute(0, 2, 1).contiguous() def get_face(self, image_pil): orig_w, orig_h = image_pil.size x = self.transform(image_pil).unsqueeze(0).to(self.device) with torch.no_grad(): out = self.model_seg(x, self.embedding, l_mask=self.l_mask4seg) out = out.argmax(dim=1, keepdim=True).float() out = F.interpolate(out, size=(orig_h, orig_w), mode="nearest") mask = (out.squeeze(0).squeeze(0).cpu().numpy() > 0).astype(np.uint8) * 255 rgb = np.array(image_pil, dtype=np.uint8) rgba = np.dstack([rgb, mask]) return rgba