import torch import torch.nn as nn import sys sys.path.append("..") from prompts.imagenet_template import openai_imagenet_template, sub_imagenet_template from mmseg.models.segmentors import BaseSegmentor from mmseg.models.data_preprocessor import SegDataPreProcessor from mmengine.structures import PixelData from mmseg.registry import MODELS from torchvision import transforms import torch.nn.functional as F from einops import rearrange from open_clip import create_model, tokenizer from segment_anything import sam_model_registry from myutils import UnNormalize @MODELS.register_module() class ProxyCLIPSegmentation(BaseSegmentor): def __init__(self, clip_type, model_type, vfm_model, name_path, checkpoint=None, device=torch.device('cuda'), prob_thd=0.0, logit_scale=40, beta=1.2, gamma=3.0, slide_stride=112, slide_crop=336): data_preprocessor = SegDataPreProcessor( mean=[122.771, 116.746, 104.094], std=[68.501, 66.632, 70.323], bgr_to_rgb=True ) super().__init__(data_preprocessor=data_preprocessor) self.clip = create_model(model_type, pretrained=clip_type, precision='fp16') self.clip.eval().to(device) self.tokenizer = tokenizer.tokenize self.vfm_model = vfm_model sam_ckpts = { "sam-B": "/mnt/SSD8T/home/wjj/code/ProxyCLIP/sam_ckpts/sam_vit_b_01ec64.pth", "sam-L": "/mnt/SSD8T/home/wjj/code/ProxyCLIP/sam_ckpts/sam_vit_l_0b3195.pth", } dinov2_ckpts = { "dinov2-L": "dinov2_vitl14_reg", "dinov2-B": "dinov2_vitb14_reg", "dinov2-B-noreg": "dinov2_vitb14", "dinov2-L-noreg": "dinov2_vitl14", } dino_ckpts = { "dino-B-8": "dino_vitb8", "dino-B-16": "dino_vitb16", } vfm = None if vfm_model.startswith("dinov2"): if vfm_model in dinov2_ckpts: model_name = dinov2_ckpts[vfm_model] hub_path = '/mnt/SSD8T/home/wjj/.cache/torch/hub/facebookresearch_dinov2_main' try: vfm = torch.hub.load(hub_path, model_name, source='local').half() except Exception as e: raise RuntimeError(f"Failed to load DINOv2 model '{vfm_model}': {e}") else: raise NotImplementedError(f"VLM model '{vfm_model}' not supported under DINOv2 category.") elif vfm_model.startswith("dino"): if vfm_model in dino_ckpts: model_name = dino_ckpts[vfm_model] hub_path = '/mnt/SSD8T/home/wjj/.cache/torch/hub/facebookresearch_dino_main' try: vfm = torch.hub.load(hub_path, model_name, source='local').half() except Exception as e: raise RuntimeError(f"Failed to load DINO model '{vfm_model}': {e}") else: raise NotImplementedError(f"VLM model '{vfm_model}' not supported under DINO category.") elif vfm_model.startswith("sam"): if vfm_model in sam_ckpts: vit_type = "vit_b" if "B" in vfm_model else "vit_l" checkpoint_path = sam_ckpts[vfm_model] try: vfm = sam_model_registry[vit_type](checkpoint=checkpoint_path).half() except Exception as e: raise RuntimeError(f"Failed to load SAM model '{vfm_model}' with checkpoint '{checkpoint_path}': {e}") else: # 为了向后兼容，如果只传入 'sam'，默认使用 sam-B if vfm_model == 'sam': vfm = sam_model_registry["vit_b"](checkpoint=sam_ckpts["sam-B"]).half() else: raise NotImplementedError(f"VLM model '{vfm_model}' not supported under SAM category.") else: raise NotImplementedError(f"VLM model '{vfm_model}' not supported.") for p in vfm.parameters(): p.requires_grad = False self.vfm = vfm.eval().to(device) self.unnorm = UnNormalize([0.48145466, 0.4578275, 0.40821073], [0.26862954, 0.26130258, 0.27577711]) self.norm = transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) query_words, self.query_idx = get_cls_idx(name_path) self.num_queries = len(query_words) self.num_classes = max(self.query_idx) + 1 self.query_idx = torch.Tensor(self.query_idx).to(torch.int64).to(device) query_features = [] with torch.no_grad(): for qw in query_words: query = self.tokenizer([temp(qw) for temp in openai_imagenet_template]).to(device) feature = self.clip.encode_text(query) feature /= feature.norm(dim=-1, keepdim=True) feature = feature.mean(dim=0) feature /= feature.norm() query_features.append(feature.unsqueeze(0)) self.query_features = torch.cat(query_features, dim=0).detach() self.dtype = self.query_features.dtype self.logit_scale = logit_scale self.prob_thd = prob_thd self.slide_stride = slide_stride self.slide_crop = slide_crop self.beta = beta self.gamma = gamma @torch.no_grad() def forward_feature(self, img, logit_size=None): if type(img) == list: img = img[0] clip_token_size = img.shape[-2] // self.clip.visual.patch_size[0], img.shape[-1] // self.clip.visual.patch_size[1] imgs_norm = [self.norm(self.unnorm(img[i])) for i in range(len(img))] imgs_norm = torch.stack(imgs_norm, dim=0) imgs_norm = imgs_norm.half() if self.vfm_model.startswith('sam'): patch_size = self.vfm.image_encoder.patch_embed.proj.kernel_size imgs_norm = F.interpolate(imgs_norm, size=(1024, 1024), mode='bilinear', align_corners=False) I, J = imgs_norm.shape[-2] // patch_size[0], imgs_norm.shape[-2] // patch_size[1] ex_feats = self.vfm.image_encoder(imgs_norm) elif self.vfm_model.startswith('dino') and not self.vfm_model.startswith('dinov2'): feat_out = {} def hook_fn_forward_qkv(module, input, output): feat_out["qkv"] = output self.vfm._modules["blocks"][-1]._modules["attn"]._modules["qkv"].register_forward_hook( hook_fn_forward_qkv) # Forward pass in the model feat = self.vfm.get_intermediate_layers(imgs_norm)[0] nb_im = feat.shape[0] # Batch size nb_tokens = feat.shape[1] # Number of tokens nh = self.vfm.blocks[0].attn.num_heads # Number of heads qkv = ( feat_out["qkv"] .reshape(nb_im, nb_tokens, 3, nh, -1 // nh) .permute(2, 0, 3, 1, 4) ) q, k, v = qkv[0], qkv[1], qkv[2] k = k.transpose(1, 2).reshape(nb_im, nb_tokens, -1)[:, 1:, :] q = q.transpose(1, 2).reshape(nb_im, nb_tokens, -1)[:, 1:, :] v = v.transpose(1, 2).reshape(nb_im, nb_tokens, -1)[:, 1:, :] patch_size = self.vfm.patch_embed.patch_size I, J = imgs_norm[0].shape[-2] // patch_size, imgs_norm[0].shape[-2] // patch_size # ex_feats = q.reshape(nb_im, I, J, -1).permute(0, 3, 1, 2) # ex_feats = k.reshape(nb_im, I, J, -1).permute(0, 3, 1, 2) # ex_feats = v.reshape(nb_im, I, J, -1).permute(0, 3, 1, 2) ex_feats = feat[:, 1:, :].reshape(nb_im, I, J, -1).permute(0, 3, 1, 2) elif self.vfm_model.startswith('dinov2'): patch_size = self.vfm.patch_embed.patch_size I, J = imgs_norm.shape[-2] // patch_size[0], imgs_norm.shape[-2] // patch_size[1] ex_feats = self.vfm.get_intermediate_layers(imgs_norm, reshape=True)[0] elif self.vfm_model == 'mae': patch_size = self.vfm.patch_embed.patch_size imgs_norm = F.interpolate(imgs_norm, size=(self.slide_crop, self.slide_crop), mode='bilinear', align_corners=False) I, J = imgs_norm.shape[-2] // patch_size[0], imgs_norm.shape[-2] // patch_size[1] image_feat = self.vfm.forward_features(imgs_norm) ex_feats = rearrange(image_feat, 'b (h w) c -> b c h w', h=I, w=J) else: I, J = clip_token_size ex_feats = None image_features = self.clip.encode_image(img.half(), external_feats=ex_feats, beta=self.beta, gamma=self.gamma) image_features /= image_features.norm(dim=-1, keepdim=True) logits = image_features @ self.query_features.T logits = logits.permute(0, 2, 1).reshape(-1, logits.shape[-1], I, J) if logit_size == None: logits = nn.functional.interpolate(logits, size=img.shape[-2:], mode='bilinear') else: logits = nn.functional.interpolate(logits, size=logit_size, mode='bilinear') return logits def forward_slide(self, img, img_metas, stride=112, crop_size=224): """Inference by sliding-window with overlap. If h_crop > h_img or w_crop > w_img, the small patch will be used to decode without padding. """ if type(img) == list: img = img[0].unsqueeze(0) if type(stride) == int: stride = (stride, stride) if type(crop_size) == int: crop_size = (crop_size, crop_size) h_stride, w_stride = stride h_crop, w_crop = crop_size batch_size, _, h_img, w_img = img.shape out_channels = self.num_queries h_grids = max(h_img - h_crop + h_stride - 1, 0) // h_stride + 1 w_grids = max(w_img - w_crop + w_stride - 1, 0) // w_stride + 1 preds = img.new_zeros((batch_size, out_channels, h_img, w_img)) count_mat = img.new_zeros((batch_size, 1, h_img, w_img)) for h_idx in range(h_grids): for w_idx in range(w_grids): y1 = h_idx * h_stride x1 = w_idx * w_stride y2 = min(y1 + h_crop, h_img) x2 = min(x1 + w_crop, w_img) y1 = max(y2 - h_crop, 0) x1 = max(x2 - w_crop, 0) crop_img = img[:, :, y1:y2, x1:x2] # pad image when (image_size % patch_size != 0) H, W = crop_img.shape[2:] # original image shape pad = self.compute_padsize(H, W, 56) if any(pad): crop_img = nn.functional.pad(crop_img, pad) # zero padding crop_seg_logit = self.forward_feature(crop_img).detach() torch.cuda.empty_cache() # mask cutting for padded image if any(pad): l, t = pad[0], pad[2] crop_seg_logit = crop_seg_logit[:, :, t:t + H, l:l + W] preds += nn.functional.pad(crop_seg_logit, (int(x1), int(preds.shape[3] - x2), int(y1), int(preds.shape[2] - y2))) count_mat[:, :, y1:y2, x1:x2] += 1 assert (count_mat == 0).sum() == 0 preds = preds / count_mat img_size = img_metas[0]['ori_shape'][:2] logits = nn.functional.interpolate(preds, size=img_size, mode='bilinear') return logits def predict(self, inputs, data_samples): if data_samples is not None: batch_img_metas = [ data_sample.metainfo for data_sample in data_samples ] else: batch_img_metas = [ dict( ori_shape=inputs.shape[2:], img_shape=inputs.shape[2:], pad_shape=inputs.shape[2:], padding_size=[0, 0, 0, 0]) ] * inputs.shape[0] if self.slide_crop > 0: seg_logits = self.forward_slide(inputs, batch_img_metas, self.slide_stride, self.slide_crop) else: seg_logits = self.forward_feature(inputs, batch_img_metas[0]['ori_shape']) return self.postprocess_result(seg_logits, data_samples) def postprocess_result(self, seg_logits, data_samples): batch_size = seg_logits.shape[0] for i in range(batch_size): seg_logits = seg_logits[i] * self.logit_scale seg_logits = seg_logits.softmax(0) # n_queries * w * h num_cls, num_queries = max(self.query_idx) + 1, len(self.query_idx) if num_cls != num_queries: seg_logits = seg_logits.unsqueeze(0) cls_index = nn.functional.one_hot(self.query_idx) cls_index = cls_index.T.view(num_cls, num_queries, 1, 1) seg_logits = (seg_logits * cls_index).max(1)[0] seg_pred = seg_logits.argmax(0, keepdim=True) seg_pred[seg_logits.max(0, keepdim=True)[0] < self.prob_thd] = 0 if data_samples is None: return seg_pred else: data_samples[i].set_data({ 'seg_logits': PixelData(**{'data': seg_logits}), 'pred_sem_seg': PixelData(**{'data': seg_pred}) }) return data_samples def compute_padsize(self, H: int, W: int, patch_size: int): l, r, t, b = 0, 0, 0, 0 if W % patch_size: lr = patch_size - (W % patch_size) l = lr // 2 r = lr - l if H % patch_size: tb = patch_size - (H % patch_size) t = tb // 2 b = tb - t return l, r, t, b def _forward(data_samples): """ """ def inference(self, img, batch_img_metas): """ """ def encode_decode(self, inputs, batch_img_metas): """ """ def extract_feat(self, inputs): """ """ def loss(self, inputs, data_samples): """ """ def get_cls_idx(path): with open(path, 'r') as f: name_sets = f.readlines() num_cls = len(name_sets) class_names, class_indices = [], [] for idx in range(num_cls): names_i = name_sets[idx].split('; ') class_names += names_i class_indices += [idx for _ in range(len(names_i))] class_names = [item.replace('\n', '') for item in class_names] return class_names, class_indices