| """ |
| SAM-GSC 消融实验训练代码 |
| |
| 使用 SAM image encoder 的 self-attention 代替 SD attention 来 refine DINO 相似度矩阵。 |
| 实时计算 SAM attention(不预提取)。 |
| """ |
|
|
| from typing import List, Tuple |
| import torch |
| import torch.nn.functional as F |
| import torch.nn as nn |
| from torchvision.ops import roi_align |
|
|
| from src.segment_anything import sam_model_registry |
| from src.segment_anything.modeling.image_encoder import Attention as SAMAttention |
|
|
|
|
| |
|
|
| class SAMAttentionExtractor(nn.Module): |
| """ |
| 从 SAM image encoder 提取 global attention layers 的 attention map |
| """ |
| def __init__(self, sam_checkpoint: str, model_type: str = "vit_l", |
| attention_layer_indices: List[int] = None, device: str = "cuda"): |
| super().__init__() |
| |
| |
| sam = sam_model_registry[model_type](checkpoint=sam_checkpoint) |
| self.image_encoder = sam.image_encoder.to(device).half().eval() |
| |
| |
| for p in self.image_encoder.parameters(): |
| p.requires_grad = False |
| |
| |
| self.global_attn_layer_indices = [] |
| for i, blk in enumerate(self.image_encoder.blocks): |
| if blk.window_size == 0: |
| self.global_attn_layer_indices.append(i) |
| |
| |
| if attention_layer_indices is None: |
| self.attention_layers = self.global_attn_layer_indices[-2:] |
| else: |
| self.attention_layers = [self.global_attn_layer_indices[i] for i in attention_layer_indices] |
| |
| |
| self._patch_attention_modules() |
| |
| def _patch_attention_modules(self): |
| """Patch attention modules to return attention weights""" |
| for layer_idx in self.attention_layers: |
| block = self.image_encoder.blocks[layer_idx] |
| original_attn = block.attn |
| block.attn = SAMAttentionWithOutput(original_attn) |
| |
| @torch.no_grad() |
| def forward(self, x: torch.Tensor) -> torch.Tensor: |
| """ |
| 提取 SAM attention map |
| |
| Args: |
| x: 输入图像 (B, 3, H, W),已经过预处理 |
| |
| Returns: |
| attention: 聚合的 attention map (B, HW, HW) |
| """ |
| |
| x = self.image_encoder.patch_embed(x) |
| _, h, w, _ = x.shape |
| |
| |
| if self.image_encoder.pos_embed is not None: |
| if (h, w) == self.image_encoder.grid_size: |
| x = x + self.image_encoder.pos_embed |
| else: |
| x = x + self.image_encoder.rescale_positional_embedding(out_size=(h, w), dtype=x.dtype) |
| |
| |
| attentions = [] |
| |
| for i, blk in enumerate(self.image_encoder.blocks): |
| if i in self.attention_layers: |
| |
| shortcut = x |
| x_normed = blk.norm1(x) |
| x_attn, attn = blk.attn(x_normed, return_attention=True) |
| x = shortcut + x_attn |
| x = x + blk.mlp(blk.norm2(x)) |
| attentions.append(attn) |
| else: |
| x = blk(x) |
| |
| |
| if len(attentions) > 0: |
| attn_stack = torch.stack(attentions, dim=0) |
| attn_aggregated = attn_stack.mean(dim=(0, 2)) |
| else: |
| B = x.shape[0] |
| HW = h * w |
| attn_aggregated = torch.eye(HW, device=x.device, dtype=x.dtype).unsqueeze(0).expand(B, -1, -1) |
| |
| return attn_aggregated |
|
|
|
|
| class SAMAttentionWithOutput(nn.Module): |
| """修改 SAM Attention 模块以返回 attention weights""" |
| |
| def __init__(self, original_attn: SAMAttention): |
| super().__init__() |
| self.num_heads = original_attn.num_heads |
| self.scale = original_attn.scale |
| self.qkv = original_attn.qkv |
| self.proj = original_attn.proj |
| self.use_rel_pos = original_attn.use_rel_pos |
| if self.use_rel_pos: |
| self.rel_pos_h = original_attn.rel_pos_h |
| self.rel_pos_w = original_attn.rel_pos_w |
| |
| def forward(self, x: torch.Tensor, return_attention: bool = False): |
| B, H, W, _ = x.shape |
| qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4) |
| q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0) |
|
|
| attn = (q * self.scale) @ k.transpose(-2, -1) |
|
|
| if self.use_rel_pos: |
| from src.segment_anything.modeling.image_encoder import add_decomposed_rel_pos |
| attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W)) |
|
|
| attn = attn.softmax(dim=-1) |
| |
| x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1) |
| x = self.proj(x) |
| |
| if return_attention: |
| |
| attn_output = attn.view(B, self.num_heads, H * W, H * W) |
| return x, attn_output |
| return x |
|
|
|
|
| |
|
|
| class DeCLIPWithREPAProjector(nn.Module): |
| """与 declip_plus.py 保持一致的模型包装器""" |
| |
| def __init__(self, declip_model, clip_dim=768, hidden_dim=1024, vfm_dim=768, args=None): |
| super().__init__() |
| self.model = declip_model |
| self.repa_layer_idx = args.repa_layer_idx |
| self.projector = nn.Sequential( |
| nn.Linear(clip_dim, hidden_dim), |
| nn.SiLU(), |
| nn.Linear(hidden_dim, hidden_dim), |
| nn.SiLU(), |
| nn.Linear(hidden_dim, vfm_dim) |
| ) |
| self.initialize_projector_weights() |
| self.logit_scale = self.model.logit_scale |
|
|
| def initialize_projector_weights(self): |
| for module in self.projector.modules(): |
| if isinstance(module, nn.Linear): |
| nn.init.xavier_uniform_(module.weight) |
| if module.bias is not None: |
| nn.init.constant_(module.bias, 0) |
|
|
| def encode_image(self, *args, **kwargs): |
| return self.model.encode_image(*args, **kwargs) |
|
|
| def encode_text(self, *args, **kwargs): |
| return self.model.encode_text(*args, **kwargs) |
|
|
| def encode_dense(self, *args, **kwargs): |
| return self.model.encode_dense(*args, **kwargs) |
|
|
| def encode_pseudo_boxes(self, images, rois_list, normalize=False, mode="qq", size=(1, 1)): |
| student_roi_features, context, intermediate_layer_output = self.model.encode_pseudo_boxes( |
| images, rois_list, normalize=normalize, mode=mode, size=size, |
| get_intermediate_layer=[self.repa_layer_idx] |
| ) |
| alpha = 0.3 |
| residual = intermediate_layer_output[0] |
| intermediate_layer_output = self.projector(intermediate_layer_output[0]) |
| intermediate_layer_output = alpha * residual + intermediate_layer_output |
| return student_roi_features, context, intermediate_layer_output |
|
|
| def encode_masks(self, *args, **kwargs): |
| return self.model.encode_masks(*args, **kwargs) |
|
|
| def train(self, mode=True): |
| self.model.train(mode) |
| self.training = self.model.training |
| return self |
|
|
| def lock_image_tower(self, *args, **kwargs): |
| return self.model.lock_image_tower(*args, **kwargs) |
|
|
| def lock_text_tower(self, *args, **kwargs): |
| return self.model.lock_text_tower(*args, **kwargs) |
|
|
| @torch.jit.ignore |
| def set_grad_checkpointing(self, enable=True): |
| self.model.set_grad_checkpointing(enable) |
|
|
| @torch.jit.ignore |
| def no_weight_decay(self): |
| return self.model.no_weight_decay() |
|
|
|
|
| class DeCLIP_SAM_GSC: |
| """ |
| SAM-GSC 消融实验:使用 SAM attention 代替 SD attention |
| """ |
| |
| def __init__(self, sam_extractor: SAMAttentionExtractor): |
| self.sam_extractor = sam_extractor |
|
|
| def __call__(self, batch, student, teacher, vfm_model, args): |
| losses = {} |
| context_weight = args.loss_context_weight |
| content_weight = args.loss_content_weight |
| region_weight = args.loss_region_weight |
| need_repa = args.repa_layer_idx != -1 |
| |
| if args.distributed: |
| student = student.module |
| |
| dtype_map = {"bf16": torch.bfloat16, "amp": torch.float16} |
| input_dtype = dtype_map.get(args.precision, torch.float32) |
| |
| |
| images, normed_boxes, image_crops, vfm_image, sam_image = prepare_inputs_sam(batch, args.device, input_dtype) |
| |
| |
| with torch.no_grad(): |
| sam_attn = self.sam_extractor(sam_image) |
| |
| loss_ensemble = self.intra_image_distill( |
| student, teacher, vfm_model, images, normed_boxes, image_crops, vfm_image, sam_attn, args |
| ) |
| |
| loss_context, loss_content, loss_region = loss_ensemble[0], loss_ensemble[1], loss_ensemble[2] |
| losses.update({"loss_context": loss_context * context_weight}) |
| losses.update({"loss_content": loss_content * content_weight}) |
| losses.update({"loss_region": loss_region * region_weight}) |
| |
| if need_repa: |
| loss_repa = loss_ensemble[2] |
| losses.update({"loss_repa": loss_repa}) |
| |
| return losses, len(images) |
|
|
| def intra_image_distill(self, student, teacher, vfm_model, images, normed_boxes, image_crops, vfm_image, sam_attn, args): |
| need_repa = args.repa_layer_idx != -1 |
| roi_size = (3, 3) |
| B = images.shape[0] |
| rois_list = [] |
| crops_list = [] |
|
|
| for bboxes_per_image, crops_per_image in zip(normed_boxes, image_crops): |
| valid = bboxes_per_image[:, -1] > 0.5 |
| rois_list.append(bboxes_per_image[valid, :4]) |
| crops_list.append(crops_per_image[valid]) |
|
|
| image_crops = torch.cat(crops_list) |
| |
| if need_repa: |
| student_roi_features, context, intermediate_layer_output = student.encode_pseudo_boxes( |
| images, rois_list, normalize=True, mode=args.mode, size=roi_size |
| ) |
| else: |
| student_roi_features, context = student.encode_pseudo_boxes( |
| images, rois_list, normalize=True, mode=args.mode, size=roi_size |
| ) |
|
|
| with torch.no_grad(): |
| teacher_crop_features = teacher.encode_image(image_crops, normalize=True) |
| intra_vfm_feats = extract_vfm_features(vfm_model, vfm_image, args) |
| vfm_roi_features = extract_roi_features(intra_vfm_feats, rois_list, normalize=True) |
| intra_vfm_feats = F.normalize(intra_vfm_feats, dim=1).flatten(start_dim=-2) |
| intra_vfm_corr = torch.einsum('bci,bcj->bij', intra_vfm_feats, intra_vfm_feats) |
| |
| |
| refined_intra_vfm_corr = refine_dino_with_sam(intra_vfm_corr, sam_attn, args.sd_refine_weight) |
|
|
| student_intra_corr = compute_student_intra_image_similarity(images.shape[0], context, args) |
| loss_context = context_loss(student_intra_corr, refined_intra_vfm_corr, teacher_temp=0.8) |
| loss_content = soft_content_distill_loss(student_roi_features, teacher_crop_features) |
| loss_region = region_scd_loss(student_roi_features, vfm_roi_features) |
| |
| if need_repa: |
| loss_repa = repa_loss(intermediate_layer_output, intra_vfm_feats) |
| return loss_context, loss_content, loss_repa |
| else: |
| return loss_context, loss_content, loss_region |
|
|
|
|
| |
|
|
| def refine_dino_with_sam(dino_corr: torch.Tensor, sam_attn: torch.Tensor, refine_weight: float) -> torch.Tensor: |
| """使用 SAM attention refine DINO 相似度矩阵""" |
| |
| B_dino, HW_dino, _ = dino_corr.shape |
| B_sam, HW_sam, _ = sam_attn.shape |
| |
| if HW_dino != HW_sam: |
| sam_attn = resize_attention(sam_attn, int(HW_dino ** 0.5)) |
| |
| residual = dino_corr |
| dino_corr_propagated = torch.bmm(sam_attn, dino_corr) |
| dino_corr_refined = dino_corr_propagated * refine_weight + residual * (1 - refine_weight) |
| |
| |
| bs, hw, _ = dino_corr_refined.shape |
| device = dino_corr_refined.device |
| eye = torch.eye(hw, dtype=dino_corr_refined.dtype, device=device).unsqueeze(0).expand(bs, -1, -1) |
| dino_corr_refined = dino_corr_refined * (1 - eye) + eye |
| |
| return dino_corr_refined |
|
|
|
|
| def resize_attention(attn: torch.Tensor, target_size: int) -> torch.Tensor: |
| """调整 attention 矩阵尺寸""" |
| B, N, _ = attn.shape |
| current_size = int(N ** 0.5) |
| |
| if current_size == target_size: |
| return attn |
| |
| |
| attn = attn.view(B, current_size, current_size, current_size, current_size) |
| attn = attn.permute(0, 1, 3, 2, 4).contiguous() |
| attn = attn.view(B, current_size * current_size, current_size, current_size) |
| attn = F.interpolate(attn, size=(target_size, target_size), mode='bilinear', align_corners=False) |
| attn = attn.view(B, current_size, current_size, target_size * target_size) |
| attn = attn.permute(0, 3, 1, 2).contiguous() |
| attn = F.interpolate(attn, size=(target_size, target_size), mode='bilinear', align_corners=False) |
| attn = attn.view(B, target_size * target_size, target_size * target_size) |
| attn = F.softmax(attn, dim=-1) |
| |
| return attn |
|
|
|
|
| def prepare_inputs_sam(batch, device, dtype): |
| """准备 SAM-GSC 的输入(包含 SAM 图像)""" |
| images, normed_boxes, image_crops, vfm_image, sam_image = batch |
| images = images.to(device=device, dtype=dtype, non_blocking=True) |
| normed_boxes = normed_boxes.to(device=device, dtype=dtype, non_blocking=True) |
| image_crops = image_crops.to(device=device, dtype=dtype, non_blocking=True) |
| vfm_image = vfm_image.to(device=device, dtype=dtype, non_blocking=True) |
| sam_image = sam_image.to(device=device, dtype=dtype, non_blocking=True) |
| return images, normed_boxes, image_crops, vfm_image, sam_image |
|
|
|
|
| def extract_vfm_features(vfm_model, image, args): |
| """从 VFM 模型提取特征""" |
| if "dinov2" in args.use_vfm or "sd_dino" in args.use_vfm or "sam_dino" in args.use_vfm: |
| vfm_feats = vfm_model.get_intermediate_layers(image, reshape=True)[0] |
| elif 'sam' in args.use_vfm: |
| vfm_feats = vfm_model.image_encoder(image) |
| elif 'dino' in args.use_vfm: |
| feat = vfm_model.get_intermediate_layers(image)[0] |
| nb_im = feat.shape[0] |
| patch_size = vfm_model.patch_embed.patch_size |
| I, J = image[0].shape[-2] // patch_size, image[0].shape[-2] // patch_size |
| vfm_feats = feat[:, 1:, :].reshape(nb_im, I, J, -1).permute(0, 3, 1, 2) |
| else: |
| raise NotImplementedError(f"VFM mode {args.use_vfm} is not implemented.") |
| return vfm_feats |
|
|
|
|
| def extract_roi_features(x, normed_boxes, size=(3, 3), normalize=False): |
| """提取 ROI 特征""" |
| def _denormalize_boxes(normed_boxes, x): |
| h, w = x.shape[-2:] |
| denormed_boxes = [] |
| for boxes in normed_boxes: |
| new_boxes = boxes.clone() |
| new_boxes[:, [0, 2]] *= w |
| new_boxes[:, [1, 3]] *= h |
| denormed_boxes.append(new_boxes) |
| return denormed_boxes |
| |
| if size == (1, 1): |
| roi_feats = roi_align(x, _denormalize_boxes(normed_boxes, x), size, 1.0, -1, True)[..., 0, 0] |
| else: |
| roi_feats = roi_align(x, _denormalize_boxes(normed_boxes, x), size, 1.0, -1, True).flatten(start_dim=-2) |
| |
| if normalize: |
| roi_feats = F.normalize(roi_feats, dim=1) |
| return roi_feats |
|
|
|
|
| def compute_student_intra_image_similarity(B, context, args): |
| """计算学生模型的图像内相似度""" |
| N, _ = context[0].shape[1:] if isinstance(context, tuple) else context.shape[1:] |
| |
| if args.mode in ["qq_vfm_distill", "kk_vfm_distill", "vv_vfm_distill", "sanity_check"]: |
| context = context.transpose(0, 1).contiguous().view(N, B, -1).transpose(0, 1) |
| context = F.normalize(context, dim=-1).transpose(-2, -1) |
| student_context_similarity = torch.einsum("b c m, b c n -> b m n", context, context) |
| elif args.mode == "csa_vfm_distill": |
| q_feature, k_feature = context |
| q_feature = q_feature.transpose(0, 1).contiguous().view(N, B, -1).transpose(0, 1) |
| k_feature = k_feature.transpose(0, 1).contiguous().view(N, B, -1).transpose(0, 1) |
| q_feature = F.normalize(q_feature, dim=-1).transpose(-2, -1) |
| k_feature = F.normalize(k_feature, dim=-1).transpose(-2, -1) |
| student_context_similarity = ( |
| torch.einsum("b c m, b c n -> b m n", q_feature, q_feature) + |
| torch.einsum("b c m, b c n -> b m n", k_feature, k_feature) |
| ) / 2.0 |
| elif args.mode == "all_vfm_distill": |
| q_feature, k_feature, v_feature = context |
| features = [q_feature, k_feature, v_feature] |
| similarities = [] |
| for feature in features: |
| feature = feature.transpose(0, 1).contiguous().view(N, B, -1).transpose(0, 1) |
| feature = F.normalize(feature, dim=-1).transpose(-2, -1) |
| similarities.append(torch.einsum("b c m, b c n -> b m n", feature, feature)) |
| student_context_similarity = sum(similarities) / len(features) |
| else: |
| raise NotImplementedError(f"Mode '{args.mode}' is not implemented.") |
| |
| return student_context_similarity |
|
|
|
|
| def context_loss(student_corr, teacher_corr, teacher_temp=1.0, student_temp=1.0): |
| """Context distillation loss""" |
| student_log_prob = F.log_softmax(student_corr / student_temp, dim=-1) |
| with torch.no_grad(): |
| teacher_prob = F.softmax(teacher_corr / teacher_temp, dim=-1) |
| kl_loss = F.kl_div(student_log_prob, teacher_prob, reduction='batchmean') * (teacher_temp * student_temp) |
| return kl_loss |
|
|
|
|
| def soft_content_distill_loss(student_roi_features, teacher_crop_features, T=1.0): |
| """Content distillation loss""" |
| sim = torch.einsum('bpc,bc->bp', student_roi_features, teacher_crop_features) |
| weights = F.softmax(sim / T, dim=1) |
| weighted_student = (student_roi_features * weights.unsqueeze(-1)).sum(dim=1) |
| weighted_student = F.normalize(weighted_student, dim=-1) |
| cosine_similarity = (weighted_student * teacher_crop_features).sum(dim=-1) |
| loss = 1.0 - cosine_similarity.mean() |
| return loss |
|
|
|
|
| def region_scd_loss(student_roi_features, intra_vfm_roi_feats, T_teacher=1.0, T_student=1.0): |
| """Region correlation loss""" |
| with torch.no_grad(): |
| intra_vfm_roi_feats = intra_vfm_roi_feats.transpose(-2, -1).contiguous() |
| teacher_corr = torch.einsum('bic,bjc->bij', intra_vfm_roi_feats, intra_vfm_roi_feats) / T_teacher |
| teacher_prob = F.softmax(teacher_corr, dim=-1) |
| student_corr = torch.einsum('bic,bjc->bij', student_roi_features, student_roi_features) / T_student |
| student_log_prob = F.log_softmax(student_corr, dim=-1) |
| loss = F.kl_div(student_log_prob, teacher_prob, reduction='batchmean') * (T_teacher * T_student) |
| return loss |
|
|
|
|
| def repa_loss(clip_intermediate_out, vfm_out): |
| """REPA loss""" |
| vfm_out = vfm_out.transpose(1, 2) |
| clip_intermediate_out = clip_intermediate_out[:, 1:] |
| clip_intermediate_out = F.normalize(clip_intermediate_out, dim=-1) |
| similarity = (clip_intermediate_out * vfm_out).sum(dim=-1) |
| loss = -similarity.mean() |
| return loss |
|
|
|
|
| |
|
|
| def build_sam_attention_extractor(args): |
| """构建 SAM attention 提取器""" |
| sam_ckpts = { |
| "sam-B": "/opt/tiger/xiaomoguhzz/sam_vit_b_01ec64.pth", |
| "sam-L": "/opt/tiger/xiaomoguhzz/sam_vit_l_0b3195.pth", |
| } |
| |
| |
| sam_type = getattr(args, 'sam_type', 'sam-L') |
| checkpoint = sam_ckpts.get(sam_type, sam_ckpts['sam-L']) |
| model_type = "vit_l" if "L" in sam_type else "vit_b" |
| |
| extractor = SAMAttentionExtractor( |
| sam_checkpoint=checkpoint, |
| model_type=model_type, |
| device=args.device |
| ) |
| return extractor |
|
|