from collections import OrderedDict from typing import Tuple, Union import numpy as np import torch import torch.nn.functional as F from torch import nn import loralib as lora import math import collections import torch.nn.init as init # import spconv.pytorch as spconv import sys sys.path.append('/home/aiops/wangzh/llava') from depth_anything_v2.dpt import DepthAnythingV2 class CPEconv(nn.Module): def __init__(self, in_channels, spatial_shape, kernel_size=(3, 3, 3), padding=(1, 1, 1)): super(CPEconv, self).__init__() self.in_channels = in_channels self.spatial_shape = 6 self.conv3d = nn.Conv3d(in_channels, in_channels, kernel_size=kernel_size, padding=padding,groups=in_channels) nn.init.zeros_(self.conv3d.weight) if self.conv3d.bias is not None: nn.init.zeros_(self.conv3d.bias) self.register_buffer('target_tensor_template', torch.zeros(1, in_channels, self.spatial_shape, 1, 1)) def generate_3d_coords_from_depth(self, depth_maps): # 假设 depth_maps 形状为 (B, H, W) B, H, W = depth_maps.shape z_min = depth_maps.min(dim=-1, keepdim=True)[0].min(dim=-2, keepdim=True)[0] # (B, 1, 1) z_max = depth_maps.max(dim=-1, keepdim=True)[0].max(dim=-2, keepdim=True)[0] # (B, 1, 1) z = (depth_maps - z_min) / (z_max - z_min + 1e-8) # z = depth_maps # z 坐标为深度值，形状为 (B, H, W) return z def forward(self, features, depth): #features [197,256,768] depth [256,14,14] B,h,w=depth.shape _,_,C=features.shape D = self.spatial_shape features = features[1:,:,:] features = features.permute(1,0,2) coord=self.generate_3d_coords_from_depth(depth) bnd=self.spatial_shape - 1 coord = (coord *bnd).to(torch.int64) coord = ( coord.clamp(0, bnd) # clamp into bnd ) target_tensor = self.target_tensor_template.expand(B, C, D, h, w).clone() # target_tensor = torch.zeros(B, C, D, h, w).to(device=features.device) # return 0 coord = coord.unsqueeze(1).expand(-1, C, -1, -1) # [B, C, H, W] # reshape features 以便与 coord 进行操作 features = features.view(B, h, w, C) # [B, H, W, C] features = features.permute(0, 3, 1, 2) # [B, C, H, W] features = features.unsqueeze(2).to(dtype=target_tensor.dtype) coord = coord.unsqueeze(2) # import pdb;pdb.set_trace() # scatter features into target_tensor target_tensor = target_tensor.scatter_(2, coord, features) # 2. 使用 b 的值作为下标，将 features 的值复制到目标张量的相应位置 # 3. 使用 for 循环将 features 的值复制到目标张量 # for i in range(B): # for j in range(h): # for k in range(w): # # 获取在 features 中的索引 # index = coord[i, j, k] # 从 b 中获取索引 # target_tensor[i, :,index, j, k] = features[i, j * 14 + k, :] # 复制对应的 features 值 output = self.conv3d(target_tensor).mean(dim=2) #(B,768,14,14) output = output.reshape(-1,output.size(0),output.size(1)) cls_feat = torch.zeros(1,output.size(-2), output.size(-1)).to(device=output.device,dtype=output.dtype) out_feat = torch.cat([cls_feat,output],dim=0) return out_feat class RPE(torch.nn.Module): def __init__(self, patch_num, num_heads): super(RPE, self).__init__() self.num_heads = num_heads self.pos_bnd = patch_num self.rpe_num = 2 * self.pos_bnd + 1 self.rpe_table = torch.nn.Parameter(torch.zeros(3 * self.rpe_num, num_heads)) # torch.nn.init.trunc_normal_(self.rpe_table, std=0.02) def generate_3d_coords_from_depth(self,depth_maps): # 假设 depth_maps 形状为 (B, H, W) B, H, W = depth_maps.shape # 生成网格 i, j，形状为 (H, W) i, j = torch.meshgrid(torch.arange(H, device=depth_maps.device), torch.arange(W, device=depth_maps.device), indexing='ij') # 归一化 x 和 y 坐标 x = j.float() / (W - 1) # (H, W) y = i.float() / (H - 1) # (H, W) # 将 x 和 y 扩展到 (B, H, W) 以匹配 depth_maps x = x.unsqueeze(0).expand(B, -1, -1) # (B, H, W) y = y.unsqueeze(0).expand(B, -1, -1) # (B, H, W) z_min = depth_maps.min(dim=-1, keepdim=True)[0].min(dim=-2, keepdim=True)[0] # (B, 1, 1) z_max = depth_maps.max(dim=-1, keepdim=True)[0].max(dim=-2, keepdim=True)[0] # (B, 1, 1) z = (depth_maps - z_min) / (z_max - z_min + 1e-8) # z = depth_maps # z 坐标为深度值，形状为 (B, H, W) # 组合成 (B, H, W, 3) 的三维坐标 coords = torch.stack([x, y, z], dim=-1) # (B, H, W, 3) return coords def compute_relative_positions(self,absolute_coords): """ 计算相对位置编码 参数: absolute_coords: 形状为 (N, 3) 的绝对三维坐标张量 返回: 相对位置编码，形状为 (N, N, 3) """ # 确保输入是一个张量 if not isinstance(absolute_coords, torch.Tensor): raise ValueError("Input must be a PyTorch tensor.") N = absolute_coords.shape[1] relative_positions = absolute_coords.unsqueeze(2) - absolute_coords.unsqueeze(1) return relative_positions def forward(self,depth): # B,K,K,3 # import pdb;pdb.set_trace() depth=self.generate_3d_coords_from_depth(depth) depth=depth.reshape(depth.size(0),-1,depth.size(-1)) # zeros_tensor = torch.zeros(depth.size(0), 1, depth.size(-1)) # depth = torch.cat((zeros_tensor,depth), dim=1) coord=self.compute_relative_positions(depth) # 将 coord 从 [0, 1] 范围转换为 [0, width] 或 [0, height] # coord = coord.reshape(coord.size(0),-1,coord.size(-1)) # import pdb;pdb.set_trace() coord = (coord * torch.tensor([self.pos_bnd, self.pos_bnd, self.pos_bnd], device=coord.device)).round().long() idx = ( coord.clamp(-self.pos_bnd, self.pos_bnd) # clamp into bnd + self.pos_bnd # relative position to positive index + torch.arange(3, device=coord.device) * self.rpe_num # x, y, z stride ) out = self.rpe_table.index_select(0, idx.reshape(-1)) # out = out.reshape(coord.size(0) ,coord.size(1) ,coord.size(2) , -1) out = out.view(idx.shape + (-1,)).sum(3) out = out.permute(0, 3, 1, 2) # (N, K, K, H) -> (N, H, K, K) # out_new=torch.zeros(out.size(0),out.size(1),out.size(2)+1,out.size(3)+1) # out_new[:, :, 1:, 1:] = out return out class PositionEmbeddingCoordsSine(nn.Module): def __init__( self, temperature=10000, normalize=False, scale=None, pos_type="fourier", d_pos=None, d_in=3, gauss_scale=1.0, ): super().__init__() self.temperature = temperature self.normalize = normalize if scale is not None and normalize is False: raise ValueError("normalize should be True if scale is passed") if scale is None: scale = 2 * math.pi assert pos_type in ["sine", "fourier"] self.pos_type = pos_type self.scale = scale self.ln = LayerNorm(768) if pos_type == "fourier": assert d_pos is not None assert d_pos % 2 == 0 # define a gaussian matrix input_ch -> output_ch B = torch.empty((d_in, d_pos // 2)).normal_() B *= gauss_scale # self.gauss_B = nn.Parameter(B) self.register_buffer("gauss_B", B) self.d_pos = d_pos self.trans3d=nn.Conv1d(in_channels=3, out_channels=768, kernel_size=1) init.zeros_(self.trans3d.weight) if self.trans3d.bias is not None: init.zeros_(self.trans3d.bias) def get_sine_embeddings(self, xyz, num_channels, input_range): ncoords = xyz.shape[1] ndim = num_channels // xyz.shape[2] if ndim % 2 != 0: ndim -= 1 # automatically handle remainder by assiging it to the first dim rems = num_channels - (ndim * xyz.shape[2]) assert ( ndim % 2 == 0 ), f"Cannot handle odd sized ndim={ndim} where num_channels={num_channels} and xyz={xyz.shape}" final_embeds = [] prev_dim = 0 for d in range(xyz.shape[2]): cdim = ndim if rems > 0: # add remainder in increments of two to maintain even size cdim += 2 rems -= 2 if cdim != prev_dim: dim_t = torch.arange(cdim, dtype=torch.float32, device=xyz.device) dim_t = self.temperature ** (2 * (dim_t // 2) / cdim) # create batch x cdim x nccords embedding raw_pos = xyz[:, :, d] if self.scale: raw_pos *= self.scale pos = raw_pos[:, :, None] / dim_t pos = torch.stack( (pos[:, :, 0::2].sin(), pos[:, :, 1::2].cos()), dim=3 ).flatten(2) final_embeds.append(pos) prev_dim = cdim final_embeds = torch.cat(final_embeds, dim=2) return final_embeds def get_fourier_embeddings(self, xyz, num_channels=None, input_range=None): if num_channels is None: num_channels = self.gauss_B.shape[1] * 2 bsize, npoints = xyz.shape[0], xyz.shape[1] assert num_channels > 0 and num_channels % 2 == 0 d_in, max_d_out = self.gauss_B.shape[0], self.gauss_B.shape[1] d_out = num_channels // 2 # assert d_out <= max_d_out assert d_in == xyz.shape[-1] # clone coords so that shift/scale operations do not affect original tensor # import pdb;pdb.set_trace() ncoords = xyz.shape[1] if self.normalize: # xyz = shift_scale_points(xyz, src_range=input_range) pass xyz *= 2 * torch.pi xyz_proj = torch.mm(xyz.view(-1, d_in), self.gauss_B[:, :d_out]).view( bsize, npoints, d_out ) final_embeds = [xyz_proj.sin(), xyz_proj.cos()] # return batch x d_pos x npoints embedding final_embeds = torch.cat(final_embeds, dim=2) # import pdb;pdb.set_trace() # final_embeds = self.ln(final_embeds) final_embeds = F.normalize(final_embeds, p=2, dim=2) # If necessary, you can permute it back to [batch, 196, 768] return final_embeds def forward(self, depth_map, num_channels=None, input_range=None): cam_coords_tensor = self.generate_3d_coords_from_depth(depth_map) # (B, H, W, 3) # cam_coords_tensor = torch.tensor(cam_coords, dtype=torch.float16) # (B, H, W, 3) cam_coords_tensor = cam_coords_tensor.view(cam_coords_tensor.size(0), -1, 3) # (B, H*W, 3) xyz=cam_coords_tensor # import pdb;pdb.set_trace() assert xyz.ndim == 3 # xyz is batch x npoints x 3 if self.pos_type == "sine": with torch.no_grad(): return self.get_sine_embeddings(xyz, 768, input_range) elif self.pos_type == "fourier": with torch.no_grad(): return self.get_fourier_embeddings(xyz, num_channels, input_range) else: raise ValueError(f"Unknown {self.pos_type}") def positiontrans3d(self,depth_map): cam_coords_tensor = self.generate_3d_coords_from_depth(depth_map) # (B, H, W, 3) # cam_coords_tensor = torch.tensor(cam_coords, dtype=torch.float16) # (B, H, W, 3) cam_coords_tensor = cam_coords_tensor.view(cam_coords_tensor.size(0), -1, 3) # (B, H*W, 3) x=cam_coords_tensor x = x.permute(0, 2, 1) # (B, H*W, 3) -> (B, 3, H*W) x = self.trans3d(x) # 1D卷积映射 (B, 768, H*W) x = x.permute(0, 2, 1) # 转换回 (B, H*W, 768) return x def generate_3d_coords_from_depth(self, depth_maps): # 假设 depth_maps 形状为 (B, H, W) B, H, W = depth_maps.shape # 生成网格 i, j，形状为 (H, W) i, j = torch.meshgrid(torch.arange(H, device=depth_maps.device), torch.arange(W, device=depth_maps.device), indexing='ij') # 归一化 x 和 y 坐标 x = j.float() / (W - 1) # (H, W) y = i.float() / (H - 1) # (H, W) # 将 x 和 y 扩展到 (B, H, W) 以匹配 depth_maps x = x.unsqueeze(0).expand(B, -1, -1) # (B, H, W) y = y.unsqueeze(0).expand(B, -1, -1) # (B, H, W) z = depth_maps # z 坐标为深度值，形状为 (B, H, W) # 组合成 (B, H, W, 3) 的三维坐标 coords = torch.stack([x, y, z], dim=-1) # (B, H, W, 3) return coords class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1): super().__init__() # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1 self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.relu1 = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.relu2 = nn.ReLU(inplace=True) self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity() self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False) self.bn3 = nn.BatchNorm2d(planes * self.expansion) self.relu3 = nn.ReLU(inplace=True) self.downsample = None self.stride = stride if stride > 1 or inplanes != planes * Bottleneck.expansion: # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1 self.downsample = nn.Sequential(OrderedDict([ ("-1", nn.AvgPool2d(stride)), ("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)), ("1", nn.BatchNorm2d(planes * self.expansion)) ])) def forward(self, x: torch.Tensor): identity = x out = self.relu1(self.bn1(self.conv1(x))) out = self.relu2(self.bn2(self.conv2(out))) out = self.avgpool(out) out = self.bn3(self.conv3(out)) if self.downsample is not None: identity = self.downsample(x) out += identity out = self.relu3(out) return out class AttentionPool2d(nn.Module): def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None): super().__init__() self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5) self.k_proj = nn.Linear(embed_dim, embed_dim) self.q_proj = nn.Linear(embed_dim, embed_dim) self.v_proj = nn.Linear(embed_dim, embed_dim) self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim) self.num_heads = num_heads def forward(self, x): x = x.flatten(start_dim=2).permute(2, 0, 1) # NCHW -> (HW)NC x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0) # (HW+1)NC x = x + self.positional_embedding[:, None, :].to(x.dtype) # (HW+1)NC x, _ = F.multi_head_attention_forward( query=x[:1], key=x, value=x, embed_dim_to_check=x.shape[-1], num_heads=self.num_heads, q_proj_weight=self.q_proj.weight, k_proj_weight=self.k_proj.weight, v_proj_weight=self.v_proj.weight, in_proj_weight=None, in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]), bias_k=None, bias_v=None, add_zero_attn=False, dropout_p=0, out_proj_weight=self.c_proj.weight, out_proj_bias=self.c_proj.bias, use_separate_proj_weight=True, training=self.training, need_weights=False ) return x.squeeze(0) class ModifiedResNet(nn.Module): """ A ResNet class that is similar to torchvision's but contains the following changes: - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool. - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1 - The final pooling layer is a QKV attention instead of an average pool """ def __init__(self, layers, output_dim, heads, input_resolution=224, width=64): super().__init__() self.output_dim = output_dim self.input_resolution = input_resolution # the 3-layer stem self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False) self.conv1_alpha = nn.Conv2d(in_channels=1, out_channels=width // 2, kernel_size=3, stride=2, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(width // 2) self.relu1 = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(width // 2) self.relu2 = nn.ReLU(inplace=True) self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False) self.bn3 = nn.BatchNorm2d(width) self.relu3 = nn.ReLU(inplace=True) self.avgpool = nn.AvgPool2d(2) # residual layers self._inplanes = width # this is a *mutable* variable used during construction self.layer1 = self._make_layer(width, layers[0]) self.layer2 = self._make_layer(width * 2, layers[1], stride=2) self.layer3 = self._make_layer(width * 4, layers[2], stride=2) self.layer4 = self._make_layer(width * 8, layers[3], stride=2) embed_dim = width * 32 # the ResNet feature dimension self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim) def _make_layer(self, planes, blocks, stride=1): layers = [Bottleneck(self._inplanes, planes, stride)] self._inplanes = planes * Bottleneck.expansion for _ in range(1, blocks): layers.append(Bottleneck(self._inplanes, planes)) return nn.Sequential(*layers) def forward(self, x, alpha=None): def stem(x): x = self.relu1(self.bn1(self.conv1(x) + self.conv1_alpha(alpha))) x = self.relu2(self.bn2(self.conv2(x))) x = self.relu3(self.bn3(self.conv3(x))) x = self.avgpool(x) return x x = x.type(self.conv1.weight.dtype) x = stem(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.attnpool(x) return x class LayerNorm(nn.LayerNorm): """Subclass torch's LayerNorm to handle fp16.""" def forward(self, x: torch.Tensor): orig_type = x.dtype # ret = super().forward(x.type(torch.float32)) ret = super().forward(x) return ret.type(orig_type) class QuickGELU(nn.Module): def forward(self, x: torch.Tensor): return x * torch.sigmoid(1.702 * x) class Attention(nn.Module): def __init__( self, dim, num_heads=8, qkv_bias=True, scaled_cosine=False, scale_heads=False, logit_scale_max=math.log(1. / 0.01), attn_drop=0., proj_drop=0., lora_adapt=False, rank=16, patch_num=16 ): super().__init__() self.scaled_cosine = scaled_cosine self.scale_heads = scale_heads assert dim % num_heads == 0, 'dim should be divisible by num_heads' self.num_heads = num_heads self.head_dim = dim // num_heads self.scale = self.head_dim ** -0.5 self.logit_scale_max = logit_scale_max self.use_rel_pos = True # 保存相对位置编码的使用状态 self.rpe = RPE(patch_num=patch_num,num_heads=self.num_heads) self.rpe.requires_grad=True # import pdb;pdb.set_trace() # keeping in_proj in this form (instead of nn.Linear) to match weight scheme of original if lora_adapt: print("!!!!!!!!!!using lora for qkv projection!!!!!!!!!!") self.in_proj = lora.MergedLinear(dim, 3*dim, r=rank, enable_lora=[True, False, True]) else: self.in_proj = nn.Linear(dim, dim * 3) # self.in_proj_weight = nn.Parameter(torch.randn((dim * 3, dim)) * self.scale) # if qkv_bias: # self.in_proj_bias = nn.Parameter(torch.zeros(dim * 3)) # else: # self.in_proj_bias = None if self.scaled_cosine: self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1)))) else: self.logit_scale = None self.attn_drop = nn.Dropout(attn_drop) if self.scale_heads: self.head_scale = nn.Parameter(torch.ones((num_heads, 1, 1))) else: self.head_scale = None self.out_proj = nn.Linear(dim, dim) if not lora_adapt else lora.Linear(dim, dim, r=rank) self.out_drop = nn.Dropout(proj_drop) def forward(self, x, attn_mask = None,depth=None): L, N, C = x.shape q, k, v = self.in_proj(x).chunk(3, dim=-1) q = q.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1) k = k.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1) v = v.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1) if self.logit_scale is not None: attn = torch.bmm(F.normalize(q, dim=-1), F.normalize(k, dim=-1).transpose(-1, -2)) logit_scale = torch.clamp(self.logit_scale, max=self.logit_scale_max).exp() attn = attn.view(N, self.num_heads, L, L) * logit_scale attn = attn.view(-1, L, L) else: q = q * self.scale attn = torch.bmm(q, k.transpose(-2, -1)) if depth is not None: depth=depth.squeeze(1) res= self.rpe(depth) res=res.reshape(-1,res.size(-2),res.size(-1)) # import pdb;pdb.set_trace() attn[:,1:,1:]=attn[:,1:,1:]+res if attn_mask is not None: if attn_mask.dtype == torch.bool: new_attn_mask = torch.zeros_like(attn_mask, dtype=q.dtype) new_attn_mask.masked_fill_(attn_mask, float("-inf")) attn_mask = new_attn_mask attn += attn_mask attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = torch.bmm(attn, v) if self.head_scale is not None: x = x.view(N, self.num_heads, L, C) * self.head_scale x = x.view(-1, L, C) x = x.transpose(0, 1).reshape(L, N, C) x = self.out_proj(x) x = self.out_drop(x) return x, attn class CustomResidualAttentionBlock(nn.Module): def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None, lora_adapt=False, rank=16,patch_num=16): super().__init__() self.attn = Attention(d_model, n_head, lora_adapt=lora_adapt, rank=rank,patch_num=patch_num) self.ln_1 = LayerNorm(d_model) self.mlp = nn.Sequential(OrderedDict([ ("c_fc", nn.Linear(d_model, d_model * 4) if not lora_adapt else lora.Linear(d_model, d_model*4, r=rank)), ("gelu", QuickGELU()), ("c_proj", nn.Linear(d_model * 4, d_model) if not lora_adapt else lora.Linear(d_model*4, d_model, r=rank)) ])) self.ln_2 = LayerNorm(d_model) self.ln_cpe = LayerNorm(d_model) self.attn_mask = attn_mask self.cpe=CPEconv(d_model,patch_num) def attention(self, x: torch.Tensor,depth=None): self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None return self.attn(x, attn_mask=self.attn_mask,depth=depth) def forward(self, x: torch.Tensor, return_attn=False,depth=None): # import pdb;pdb.set_trace() # x ([577, 50, 1024]) # if None: shortcut=x # import pdb;pdb.set_trace() # shapes=x.shape # x= x.reshape(-1,x.size(-1)) # import pdb;pdb.set_trace() # cposi = self.cpe(x, depth).reshape(shapes) cposi = self.cpe(self.ln_cpe(x), depth) x =shortcut+cposi attn_out, attn = self.attention(self.ln_1(x),depth) x = x + attn_out x = x + self.mlp(self.ln_2(x)) if return_attn: return x, attn else: return x class ResidualAttentionBlock(nn.Module): def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None): super().__init__() self.attn = nn.MultiheadAttention(d_model, n_head) self.ln_1 = LayerNorm(d_model) self.mlp = nn.Sequential(OrderedDict([ ("c_fc", nn.Linear(d_model, d_model * 4)), ("gelu", QuickGELU()), ("c_proj", nn.Linear(d_model * 4, d_model)) ])) self.ln_2 = LayerNorm(d_model) self.attn_mask = attn_mask def attention(self, x: torch.Tensor): self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None return self.attn(x, x, x, attn_mask=self.attn_mask)[0] def forward(self, x: torch.Tensor): x = x + self.attention(self.ln_1(x)) x = x + self.mlp(self.ln_2(x)) return x class Transformer(nn.Module): def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None): super().__init__() self.width = width self.layers = layers self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)]) def forward(self, x: torch.Tensor): return self.resblocks(x) class CustomTransformer(nn.Module): def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None, lora_adapt=False, rank=16,patch_num=16): super().__init__() self.width = width self.layers = layers self.resblocks = nn.Sequential(*[CustomResidualAttentionBlock(width, heads, attn_mask, lora_adapt=lora_adapt, rank=rank,patch_num=patch_num) for _ in range(layers)]) def forward(self, x: torch.Tensor, return_attn=False,depth=None): # import pdb;pdb.set_trace() if return_attn: for i, block in enumerate(self.resblocks): if i == len(self.resblocks) - 1: return block(x, return_attn=True,depth=depth) else: x = block(x,depth=depth) assert False for block in self.resblocks: # import pdb;pdb.set_trace() x = block(x, depth=depth) # 将 depth 传递给每个模块 return x # return self.resblocks(x) # //////////////////////////////////////////////////////////////////////////////////////////// class VisionTransformer(nn.Module): def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int, lora_adapt=False, rank=16): super().__init__() self.input_resolution = input_resolution self.output_dim = output_dim self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False) self.conv1_alpha = nn.Conv2d(in_channels=1, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False) nn.init.zeros_(self.conv1_alpha.weight) scale = width ** -0.5 self.class_embedding = nn.Parameter(scale * torch.randn(width)) self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width)) # self.depth_positional_embedding = nn.Parameter(scale * torch.zeros((input_resolution // patch_size) ** 2, width)) # 用于alpha的深度编码 # self.depth_positional_embedding = PositionEmbeddingCoordsSine(temperature=10000, # normalize=True, # scale=2 * torch.pi, # pos_type="fourier", # d_pos=768, # 示例输出维度 # d_in=3, # gauss_scale=1.0 # ) # self.sine_positional_embedding = PositionEmbeddingCoordsSine(temperature=10000, # normalize=True, # scale=2 * torch.pi, # pos_type="sine", # d_pos=768, # 示例输出维度 # d_in=3, # gauss_scale=1.0 # ) # self.large_positional_embedding = PositionEmbeddingCoordsSine(temperature=10000, # normalize=True, # scale=2 * torch.pi, # pos_type="sine", # d_pos=1024, # 示例输出维度 # d_in=3, # gauss_scale=1.0 # ) # self.depth_mlp=nn.Linear(768,768) # nn.init.zeros_(self.depth_mlp.weight) # if self.depth_mlp.bias is not None: # nn.init.zeros_(self.depth_mlp.bias) self.patch_size=patch_size self.ln_pre = LayerNorm(width) self.transformer = CustomTransformer(width, layers, heads, lora_adapt=lora_adapt, rank=rank,patch_num=input_resolution // patch_size) self.ln_post = LayerNorm(width) self.proj = nn.Parameter(scale * torch.randn(width, output_dim)) def forward(self, x: torch.Tensor, alpha=None, return_attn=False,pos_embed=None): # import pdb;pdb.set_trace() x = self.conv1(x) # shape = [*, width, grid, grid] # ASSUME alpha is always not None! # import pdb;pdb.set_trace() # if pos_embed == "nodepth": # pass # else: # x = x + self.conv1_alpha(alpha) # import pdb;pdb.set_trace() x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2] x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width] x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1) # shape = [*, grid ** 2 + 1, width] # import pdb;pdb.set_trace() alpha_resized = F.adaptive_avg_pool2d(alpha, (self.input_resolution // self.patch_size, self.input_resolution // self.patch_size)) # alpha_flattened = alpha_resized.flatten(start_dim=2).permute(0, 2, 1) alpha_resized = alpha_resized.squeeze(1) # x[:, 1:] += self.depth_positional_embedding.to(x.dtype) * alpha_flattened # import pdb;pdb.set_trace() # if pos_embed == "fourier": # depth_embedding = self.depth_positional_embedding(alpha_resized) # x[:, 1:] +=self.depth_mlp(depth_embedding) # elif pos_embed == "sine": # depth_embedding = self.sine_positional_embedding(alpha_resized) # x[:, 1:] +=self.depth_mlp(depth_embedding) # elif pos_embed == "3d": # depth_embedding = self.depth_positional_embedding.positiontrans3d(alpha_resized) # x[:, 1:] +=self.depth_mlp(depth_embedding) x = x + self.positional_embedding.to(x.dtype) x = self.ln_pre(x) # import pdb;pdb.set_trace() x = x.permute(1, 0, 2) # NLD -> LND if return_attn: x, attn_last = self.transformer(x, return_attn=True,depth=alpha_resized) else: x = self.transformer(x, return_attn=False,depth=alpha_resized) x = x.permute(1, 0, 2) # LND -> NLD # x = self.ln_post(x[:, 0, :]) x = self.ln_post(x) # if self.proj is not None: # x = x @ self.proj if return_attn: return x, attn_last else: return x # ///////////////////////////////////////////////////////////////////////////////////////////////////// class CLIP(nn.Module): def __init__(self, embed_dim: int, # vision image_resolution: int, vision_layers: Union[Tuple[int, int, int, int], int], vision_width: int, vision_patch_size: int, # text context_length: int, vocab_size: int, transformer_width: int, transformer_heads: int, transformer_layers: int, lora_adapt = False, rank = 16, ): super().__init__() self.context_length = context_length if isinstance(vision_layers, (tuple, list)): vision_heads = vision_width * 32 // 64 self.visual = ModifiedResNet( layers=vision_layers, output_dim=embed_dim, heads=vision_heads, input_resolution=image_resolution, width=vision_width ) else: vision_heads = vision_width // 64 self.visual = VisionTransformer( input_resolution=image_resolution, patch_size=vision_patch_size, width=vision_width, layers=vision_layers, heads=vision_heads, output_dim=embed_dim, lora_adapt=lora_adapt, rank=rank ) self.transformer = Transformer( width=transformer_width, layers=transformer_layers, heads=transformer_heads, attn_mask=self.build_attention_mask() ) self.vocab_size = vocab_size self.token_embedding = nn.Embedding(vocab_size, transformer_width) self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width)) self.ln_final = LayerNorm(transformer_width) self.hidden_size = vision_width self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim)) self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07)) self.initialize_parameters() def initialize_parameters(self): nn.init.normal_(self.token_embedding.weight, std=0.02) nn.init.normal_(self.positional_embedding, std=0.01) if isinstance(self.visual, ModifiedResNet): if self.visual.attnpool is not None: std = self.visual.attnpool.c_proj.in_features ** -0.5 nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std) nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std) nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std) nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std) for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]: for name, param in resnet_block.named_parameters(): if name.endswith("bn3.weight"): nn.init.zeros_(param) proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5) attn_std = self.transformer.width ** -0.5 fc_std = (2 * self.transformer.width) ** -0.5 for block in self.transformer.resblocks: nn.init.normal_(block.attn.in_proj_weight, std=attn_std) nn.init.normal_(block.attn.out_proj.weight, std=proj_std) nn.init.normal_(block.mlp.c_fc.weight, std=fc_std) nn.init.normal_(block.mlp.c_proj.weight, std=proj_std) if self.text_projection is not None: nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5) def build_attention_mask(self): # lazily create causal attention mask, with full attention between the vision tokens # pytorch uses additive attention mask; fill with -inf mask = torch.empty(self.context_length, self.context_length) mask.fill_(float("-inf")) mask.triu_(1) # zero out the lower diagonal return mask @property def dtype(self): if not hasattr(self.visual, "conv1"): return self.visual.module.conv1.weight.dtype return self.visual.conv1.weight.dtype @property def device(self): return torch.device("cuda") def encode_image(self, image, alpha): assert alpha is not None return self.visual(image.type(self.dtype), alpha.type(self.dtype)) def encode_text(self, text): x = self.token_embedding(text).type(self.dtype) # [batch_size, n_ctx, d_model] x = x + self.positional_embedding.type(self.dtype) x = x.permute(1, 0, 2) # NLD -> LND x = self.transformer(x) x = x.permute(1, 0, 2) # LND -> NLD x = self.ln_final(x).type(self.dtype) # x.shape = [batch_size, n_ctx, transformer.width] # take features from the eot embedding (eot_token is the highest number in each sequence) x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection return x def our_encode_image(self,image, depth): # import pdb;pdb.set_trace() image_feature = self.visual(image, depth) # 32. 577 . 768 return image_feature def forward(self, image, text, alpha): image_features = self.encode_image(image, alpha) text_features = self.encode_text(text) # normalized features image_features = image_features / image_features.norm(dim=1, keepdim=True) text_features = text_features / text_features.norm(dim=1, keepdim=True) # cosine similarity as logits logit_scale = self.logit_scale.exp() logits_per_image = logit_scale * image_features @ text_features.t() logits_per_text = logits_per_image.t() # shape = [global_batch_size, global_batch_size] return logits_per_image, logits_per_text def convert_weights(model: nn.Module): """Convert applicable model parameters to fp16""" def _convert_weights_to_fp16(l): if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)): l.weight.data = l.weight.data.half() if l.bias is not None: l.bias.data = l.bias.data.half() if isinstance(l, nn.MultiheadAttention): for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]: tensor = getattr(l, attr) if tensor is not None: tensor.data = tensor.data.half() for name in ["text_projection", "proj"]: if hasattr(l, name): attr = getattr(l, name) if attr is not None: attr.data = attr.data.half() model.apply(_convert_weights_to_fp16) def build_model(state_dict: dict, lora_adapt=False, rank=16): vit = "visual.proj" in state_dict if vit: vision_width = state_dict["visual.conv1.weight"].shape[0] vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")]) vision_patch_size = state_dict["visual.conv1.weight"].shape[-1] grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5) image_resolution = vision_patch_size * grid_size else: counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]] vision_layers = tuple(counts) vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0] output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5) vision_patch_size = None assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0] image_resolution = output_width * 32 embed_dim = state_dict["text_projection"].shape[1] context_length = state_dict["positional_embedding"].shape[0] vocab_size = state_dict["token_embedding.weight"].shape[0] transformer_width = state_dict["ln_final.weight"].shape[0] transformer_heads = transformer_width // 64 transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith("transformer.resblocks"))) # always load lora version model = CLIP( embed_dim, image_resolution, vision_layers, vision_width, vision_patch_size, context_length, vocab_size, transformer_width, transformer_heads, transformer_layers, lora_adapt=lora_adapt, rank=rank, ) model_configs = { 'vits': {'encoder': 'vits', 'features': 64, 'out_channels': [48, 96, 192, 384]}, 'vitb': {'encoder': 'vitb', 'features': 128, 'out_channels': [96, 192, 384, 768]}, 'vitl': {'encoder': 'vitl', 'features': 256, 'out_channels': [256, 512, 1024, 1024]}, 'vitg': {'encoder': 'vitg', 'features': 384, 'out_channels': [1536, 1536, 1536, 1536]} } encoder = 'vitb' depth_model=DepthAnythingV2(**model_configs[encoder]) depth_model.load_state_dict(torch.load(f'/home/aiops/wangzh/zss/Depth-Anything-V2/checkpoints/depth_anything_v2_{encoder}.pth', map_location='cpu')) for key in ["input_resolution", "context_length", "vocab_size"]: if key in state_dict: del state_dict[key] # para_wb to linear new_state_dict = collections.OrderedDict() for k, v in state_dict.items(): if 'visual' in k: if 'in_proj_weight' in k: new_state_dict[k.replace('in_proj_weight', 'in_proj.weight')] = v elif 'in_proj_bias' in k: new_state_dict[k.replace('in_proj_bias', 'in_proj.bias')] = v else: new_state_dict[k] = v else: new_state_dict[k] = v state_dict = new_state_dict # add rgba_conv_weight if 'visual.conv1_alpha.weight' not in state_dict.keys(): # zero initialization on alpha channel rgb_weight = state_dict['visual.conv1.weight'].clone().detach() rgba_weigth = torch.zeros_like(rgb_weight)[:, 0:1, :, :] state_dict['visual.conv1_alpha.weight'] = rgba_weigth convert_weights(model) model.load_state_dict(state_dict, strict=False) return model.eval(), depth_model