first commit

model/SGraFormer.py

@@ -0,0 +1,176 @@
+## Our model was revised from https://github.com/zczcwh/PoseFormer/blob/main/common/model_poseformer.py
+
+import torch
+import torch.nn as nn
+from functools import partial
+from einops import rearrange
+from timm.models.layers import DropPath
+
+from common.opt import opts
+
+from model.Spatial_encoder import First_view_Spatial_features, Spatial_features
+from model.Temporal_encoder import Temporal__features
+
+opt = opts().parse()
+
+
+#######################################################################################################################
+class sgraformer(nn.Module):
+    def __init__(self, num_frame=9, num_joints=17, in_chans=2, embed_dim_ratio=32, depth=4,
+                 num_heads=8, mlp_ratio=2., qkv_bias=True, qk_scale=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.2, norm_layer=None):
+        """    ##########hybrid_backbone=None, representation_size=None,
+        Args:
+            num_frame (int, tuple): input frame number
+            num_joints (int, tuple): joints number
+            in_chans (int): number of input channels, 2D joints have 2 channels: (x,y)
+            embed_dim_ratio (int): embedding dimension ratio
+            depth (int): depth of transformer
+            num_heads (int): number of attention heads
+            mlp_ratio (int): ratio of mlp hidden dim to embedding dim
+            qkv_bias (bool): enable bias for qkv if True
+            qk_scale (float): override default qk scale of head_dim ** -0.5 if set
+            drop_rate (float): dropout rate
+            attn_drop_rate (float): attention dropout rate
+            drop_path_rate (float): stochastic depth rate
+            norm_layer: (nn.Module): normalization layer
+        """
+        super().__init__()
+
+        embed_dim = embed_dim_ratio * num_joints
+        out_dim = num_joints * 3  #### output dimension is num_joints * 3
+        ##Spatial_features
+        self.SF1 = First_view_Spatial_features(num_frame, num_joints, in_chans, embed_dim_ratio, depth,
+                                               num_heads, mlp_ratio, qkv_bias, qk_scale,
+                                               drop_rate, attn_drop_rate, drop_path_rate, norm_layer)
+        self.SF2 = Spatial_features(num_frame, num_joints, in_chans, embed_dim_ratio, depth,
+                                    num_heads, mlp_ratio, qkv_bias, qk_scale,
+                                    drop_rate, attn_drop_rate, drop_path_rate, norm_layer)
+        self.SF3 = Spatial_features(num_frame, num_joints, in_chans, embed_dim_ratio, depth,
+                                    num_heads, mlp_ratio, qkv_bias, qk_scale,
+                                    drop_rate, attn_drop_rate, drop_path_rate, norm_layer)
+        self.SF4 = Spatial_features(num_frame, num_joints, in_chans, embed_dim_ratio, depth,
+                                    num_heads, mlp_ratio, qkv_bias, qk_scale,
+                                    drop_rate, attn_drop_rate, drop_path_rate, norm_layer)
+
+        ## MVF
+        self.view_pos_embed = nn.Parameter(torch.zeros(1, 4, num_frame, embed_dim))
+        self.pos_drop = nn.Dropout(p=0.)
+
+        self.conv = nn.Sequential(
+            nn.BatchNorm2d(4, momentum=0.1),
+            nn.Conv2d(4, 1, kernel_size=opt.mvf_kernel, stride=1, padding=int(opt.mvf_kernel // 2), bias=False),
+            nn.ReLU(inplace=True),
+        )
+
+
+        self.conv_hop = nn.Sequential(
+            nn.BatchNorm2d(4, momentum=0.1),
+            nn.Conv2d(4, 1, kernel_size=opt.mvf_kernel, stride=1, padding=int(opt.mvf_kernel // 2), bias=False),
+            nn.ReLU(inplace=True),
+        )
+
+        self.conv_norm = nn.LayerNorm(embed_dim)
+
+        self.conv_hop_norm = nn.LayerNorm(embed_dim)
+
+
+        # Time Serial
+        self.TF = Temporal__features(num_frame, num_joints, in_chans, embed_dim_ratio, depth,
+                                        num_heads, mlp_ratio, qkv_bias, qk_scale,
+                                        drop_rate, attn_drop_rate, drop_path_rate, norm_layer)
+
+        self.head = nn.Sequential(
+            nn.LayerNorm(embed_dim),
+            nn.Linear(embed_dim, out_dim),
+        )
+
+        self.hop_w0 = nn.Parameter(torch.ones(17, 17))
+        self.hop_w1 = nn.Parameter(torch.ones(17, 17))
+        self.hop_w2 = nn.Parameter(torch.ones(17, 17))
+        self.hop_w3 = nn.Parameter(torch.ones(17, 17))
+        self.hop_w4 = nn.Parameter(torch.ones(17, 17))
+
+        self.hop_global = nn.Parameter(torch.ones(17, 17))
+
+        self.linear_hop = nn.Linear(8, 2)
+        # self.max_pool = nn.MaxPool1d(2)
+
+        self.edge_embedding = nn.Linear(17*17*4, 17*17)
+
+    def forward(self, x, hops):
+        b, f, v, j, c = x.shape
+
+        edge_embedding = self.edge_embedding(hops[0].reshape(1, -1))
+
+        ###############golbal feature#################
+        x_hop_global = x.unsqueeze(3).repeat(1, 1, 1, 17, 1, 1)
+        x_hop_global = x_hop_global - x_hop_global.permute(0, 1, 2, 4, 3, 5)
+        x_hop_global = torch.sum(x_hop_global ** 2, dim=-1)
+        hop_global = x_hop_global / torch.sum(x_hop_global, dim=-1).unsqueeze(-1)
+        hops = hops.unsqueeze(1).unsqueeze(2).repeat(1, f, v, 1, 1, 1)
+        hops1 = hop_global * hops[:, :, :, 0]
+        hops2 = hop_global * hops[:, :, :, 1]
+        hops3 = hop_global * hops[:, :, :, 2]
+        hops4 = hop_global * hops[:, :, :, 3]
+        # hops = torch.cat((hops1,hops2,hops3,hops4), dim=-1)
+        hops = torch.cat((hops1,hops2,hops3,hops4), dim=-1)
+        
+
+        x1 = x[:, :, 0]
+        x2 = x[:, :, 1]
+        x3 = x[:, :, 2]
+        x4 = x[:, :, 3]
+
+        x1 = x1.permute(0, 3, 1, 2)
+        x2 = x2.permute(0, 3, 1, 2)
+        x3 = x3.permute(0, 3, 1, 2)
+        x4 = x4.permute(0, 3, 1, 2)
+
+        hop1 = hops[:, :, 0]
+        hop2 = hops[:, :, 1]
+        hop3 = hops[:, :, 2]
+        hop4 = hops[:, :, 3]
+
+        hop1 = hop1.permute(0, 3, 1, 2)
+        hop2 = hop2.permute(0, 3, 1, 2)
+        hop3 = hop3.permute(0, 3, 1, 2)
+        hop4 = hop4.permute(0, 3, 1, 2)
+
+        ### Semantic graph transformer encoder
+        x1, hop1, MSA1, MSA2, MSA3, MSA4 = self.SF1(x1, hop1, edge_embedding)
+        x2, hop2, MSA1, MSA2, MSA3, MSA4 = self.SF2(x2, hop2, MSA1, MSA2, MSA3, MSA4, edge_embedding)
+        x3, hop3, MSA1, MSA2, MSA3, MSA4 = self.SF3(x3, hop3, MSA1, MSA2, MSA3, MSA4, edge_embedding)
+        x4, hop4, MSA1, MSA2, MSA3, MSA4 = self.SF4(x4, hop4, MSA1, MSA2, MSA3, MSA4, edge_embedding)
+        
+        ### Multi-view cross-channel fusion
+        x = torch.cat((x1.unsqueeze(1), x2.unsqueeze(1), x3.unsqueeze(1), x4.unsqueeze(1)), dim=1) + self.view_pos_embed
+        x = self.pos_drop(x)
+        x = self.conv(x).squeeze(1) + x1 + x2 + x3 + x4
+        x = self.conv_norm(x)
+
+        hop = torch.cat((hop1.unsqueeze(1), hop2.unsqueeze(1), hop3.unsqueeze(1), hop4.unsqueeze(1)), dim=1) + self.view_pos_embed
+        hop = self.pos_drop(hop)
+        # hop = self.conv_hop(hop).squeeze(1) + hop1 + hop2 + hop3 + hop4
+        # hop = self.conv_hop_norm(hop)
+        hop = self.conv(hop).squeeze(1) + hop1 + hop2 + hop3 + hop4
+        hop = self.conv_norm(hop)
+
+        x = x * hop
+
+
+        ### Temporal transformer encoder
+        x = self.TF(x)
+        
+        x = self.head(x)
+        x = x.view(b, opt.frames, j, -1)
+
+        print("=============> x.shape", x.shape)
+        return x
+
+
+# x = torch.rand((8, 27, 4, 17 , 2))
+# hops = torch.rand((8,4,17,17))
+# mvft = hmvformer(num_frame=opt.frames, num_joints=17, in_chans=2, embed_dim_ratio=32, depth=4,
+#                             num_heads=8, mlp_ratio=2., qkv_bias=True, qk_scale=None, drop_path_rate=0.1)
+# print(mvft(x, hops).shape)

model/Spatial_encoder.py

@@ -0,0 +1,343 @@
+## Our model was revised from https://github.com/zczcwh/PoseFormer/blob/main/common/model_poseformer.py
+
+import torch
+import torch.nn as nn
+from functools import partial
+from einops import rearrange
+from timm.models.layers import DropPath
+
+
+#######################################################################################################################
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+#######################################################################################################################
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+
+        self.edge_embedding = nn.Linear(17*17, 17*17)
+
+    def forward(self, x, edge_embedding):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+
+        edge_embedding = self.edge_embedding(edge_embedding)
+        edge_embedding = edge_embedding.reshape(1, 17, 17).unsqueeze(0).repeat(B, self.num_heads, 1, 1)
+        # print(edge_embedding.shape)
+
+        attn = attn + edge_embedding
+
+
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+#######################################################################################################################
+class CVA_Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.Qnorm = nn.LayerNorm(dim)
+        self.Knorm = nn.LayerNorm(dim)
+        self.Vnorm = nn.LayerNorm(dim)
+        self.QLinear = nn.Linear(dim, dim)
+        self.KLinear = nn.Linear(dim, dim)
+        self.VLinear = nn.Linear(dim, dim)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+
+        self.edge_embedding = nn.Linear(17*17, 17*17)
+
+
+
+
+    def forward(self, x, CVA_input, edge_embedding):
+        B, N, C = x.shape
+        # CVA_input = self.max_pool(CVA_input)
+        # print(CVA_input.shape)
+        q = self.QLinear(self.Qnorm(CVA_input)).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+        k = self.KLinear(self.Knorm(CVA_input)).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+        v = self.VLinear(self.Vnorm(x)).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+
+        edge_embedding = self.edge_embedding(edge_embedding)
+        edge_embedding = edge_embedding.reshape(1, 17, 17).unsqueeze(0).repeat(B, self.num_heads, 1, 1)
+
+
+
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+#######################################################################################################################
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x):
+        x = x + self.drop_path(self.attn(self.norm1(x), edge_embedding))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+#######################################################################################################################
+class Multi_Out_Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+
+        self.norm_hop1 = norm_layer(dim)
+        self.norm_hop2 = norm_layer(dim)
+        self.mlp_hop = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x, hops, edge_embedding):
+        MSA = self.drop_path(self.attn(self.norm1(x), edge_embedding))
+        MSA = self.norm_hop1(hops) * MSA
+
+        x = x + MSA
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+
+        hops = hops + MSA
+        hops = hops + self.drop_path(self.mlp_hop(self.norm_hop2(hops)))
+        
+
+        return x, hops, MSA
+
+
+#######################################################################################################################
+class Multi_In_Out_Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+
+        self.cva_attn = CVA_Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+        # self.max_pool = nn.MaxPool1d(3, stride=1, padding=1, dilation=1, return_indices=False, ceil_mode=False)
+
+
+        self.norm_hop1 = norm_layer(dim)
+        self.norm_hop2 = norm_layer(dim)
+        self.mlp_hop = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x, hops, CVA_input, edge_embedding):
+        MSA = self.drop_path(self.cva_attn(x, CVA_input, edge_embedding))
+        MSA = self.norm_hop1(hops) * MSA
+
+        x = x + MSA
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        hops = hops + MSA
+        hops = hops + self.drop_path(self.mlp_hop(self.norm_hop2(hops)))
+        return x, hops, MSA
+
+
+#######################################################################################################################
+class First_view_Spatial_features(nn.Module):
+    def __init__(self, num_frame=9, num_joints=17, in_chans=2, embed_dim_ratio=32, depth=4,
+                 num_heads=8, mlp_ratio=2., qkv_bias=True, qk_scale=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.2, norm_layer=None):
+        super().__init__()
+
+        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+
+        ### spatial patch embedding
+        self.Spatial_patch_to_embedding = nn.Linear(in_chans, embed_dim_ratio)
+        self.Spatial_pos_embed = nn.Parameter(torch.zeros(1, num_joints, embed_dim_ratio))
+
+
+        self.hop_to_embedding = nn.Linear(68, embed_dim_ratio)
+        self.hop_pos_embed = nn.Parameter(torch.zeros(1, num_joints, embed_dim_ratio))
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+
+        self.block1 = Multi_Out_Block(dim=embed_dim_ratio, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias,
+                                      qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[0],
+                                      norm_layer=norm_layer)
+        self.block2 = Multi_Out_Block(dim=embed_dim_ratio, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias,
+                                      qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[1],
+                                      norm_layer=norm_layer)
+        self.block3 = Multi_Out_Block(dim=embed_dim_ratio, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias,
+                                      qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[2],
+                                      norm_layer=norm_layer)
+        self.block4 = Multi_Out_Block(dim=embed_dim_ratio, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias,
+                                      qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[3],
+                                      norm_layer=norm_layer)
+
+        self.Spatial_norm = norm_layer(embed_dim_ratio)
+
+        self.hop_norm = norm_layer(embed_dim_ratio)
+
+    def forward(self, x, hops, edge_embedding):
+        b, _, f, p = x.shape  ##### b is batch size, f is number of frames, p is number of joints
+        x = rearrange(x, 'b c f p  -> (b f) p  c', )
+
+        x = self.Spatial_patch_to_embedding(x)
+        x += self.Spatial_pos_embed
+        x = self.pos_drop(x)
+        
+        hops = rearrange(hops, 'b c f p  -> (b f) p  c', )
+        hops = self.hop_to_embedding(hops)
+        hops += self.hop_pos_embed
+        hops = self.pos_drop(hops)
+
+
+        x, hops, MSA1 = self.block1(x, hops, edge_embedding)
+        x, hops, MSA2 = self.block2(x, hops, edge_embedding)
+        x, hops, MSA3 = self.block3(x, hops, edge_embedding)
+        x, hops, MSA4 = self.block4(x, hops, edge_embedding)
+
+        x = self.Spatial_norm(x)
+        x = rearrange(x, '(b f) w c -> b f (w c)', f=f)
+
+        hops = self.hop_norm(hops)
+        hops = rearrange(hops, '(b f) w c -> b f (w c)', f=f)
+
+        return x, hops, MSA1, MSA2, MSA3, MSA4
+
+
+#######################################################################################################################
+class Spatial_features(nn.Module):
+    def __init__(self, num_frame=9, num_joints=17, in_chans=2, embed_dim_ratio=32, depth=4,
+                 num_heads=8, mlp_ratio=2., qkv_bias=True, qk_scale=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.2, norm_layer=None):
+        super().__init__()
+
+        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+
+        ### spatial patch embedding
+        self.Spatial_patch_to_embedding = nn.Linear(in_chans, embed_dim_ratio)
+        self.Spatial_pos_embed = nn.Parameter(torch.zeros(1, num_joints, embed_dim_ratio))
+
+        self.hop_to_embedding = nn.Linear(68, embed_dim_ratio)
+        self.hop_pos_embed = nn.Parameter(torch.zeros(1, num_joints, embed_dim_ratio))
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+
+        self.block1 = Multi_In_Out_Block(
+            dim=embed_dim_ratio, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[0], norm_layer=norm_layer)
+        self.block2 = Multi_In_Out_Block(
+            dim=embed_dim_ratio, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[1], norm_layer=norm_layer)
+        self.block3 = Multi_In_Out_Block(
+            dim=embed_dim_ratio, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[2], norm_layer=norm_layer)
+        self.block4 = Multi_In_Out_Block(
+            dim=embed_dim_ratio, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[3], norm_layer=norm_layer)
+
+        self.Spatial_norm = norm_layer(embed_dim_ratio)
+
+        self.hop_norm = norm_layer(embed_dim_ratio)
+
+    def forward(self, x, hops, MSA1, MSA2, MSA3, MSA4, edge_embedding):
+        b, _, f, p = x.shape  ##### b is batch size, f is number of frames, p is number of joints
+        x = rearrange(x, 'b c f p  -> (b f) p  c', )
+
+        x = self.Spatial_patch_to_embedding(x)
+        x += self.Spatial_pos_embed
+        x = self.pos_drop(x)
+
+
+        hops = rearrange(hops, 'b c f p  -> (b f) p  c', )
+        hops = self.hop_to_embedding(hops)
+        hops += self.hop_pos_embed
+        hops = self.pos_drop(hops)
+
+
+        x, hops, MSA1 = self.block1(x, hops, MSA1, edge_embedding)
+        x, hops, MSA2 = self.block2(x, hops, MSA2, edge_embedding)
+        x, hops, MSA3 = self.block3(x, hops, MSA3, edge_embedding)
+        x, hops, MSA4 = self.block4(x, hops, MSA4, edge_embedding)
+
+
+        x = self.Spatial_norm(x)
+        x = rearrange(x, '(b f) w c -> b f (w c)', f=f)
+
+        hops = self.hop_norm(hops)
+        hops = rearrange(hops, '(b f) w c -> b f (w c)', f=f)
+
+        return x, hops, MSA1, MSA2, MSA3, MSA4

model/Temporal_encoder.py

@@ -0,0 +1,159 @@
+## Our model was revised from https://github.com/zczcwh/PoseFormer/blob/main/common/model_poseformer.py
+
+import torch
+import torch.nn as nn
+from functools import partial
+from einops import rearrange
+from timm.models.layers import DropPath
+
+from common.opt import opts
+
+opt = opts().parse()
+
+
+#######################################################################################################################
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+#######################################################################################################################
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+#######################################################################################################################
+class CVA_Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.Qnorm = nn.LayerNorm(dim)
+        self.Knorm = nn.LayerNorm(dim)
+        self.Vnorm = nn.LayerNorm(dim)
+        self.QLinear = nn.Linear(dim, dim)
+        self.KLinear = nn.Linear(dim, dim)
+        self.VLinear = nn.Linear(dim, dim)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+
+
+
+    def forward(self, x, CVA_input):
+        B, N, C = x.shape
+        q = self.QLinear(self.Qnorm(CVA_input)).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+        k = self.KLinear(self.Knorm(CVA_input)).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+        v = self.VLinear(self.Vnorm(x)).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+#######################################################################################################################
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x):
+        x = x + self.drop_path(self.attn(self.norm1(x)))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+#######################################################################################################################
+class Temporal__features(nn.Module):
+    def __init__(self, num_frame=9, num_joints=17, in_chans=2, embed_dim_ratio=32, depth=4,
+                 num_heads=8, mlp_ratio=2., qkv_bias=True, qk_scale=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.2, norm_layer=None):
+        super().__init__()
+        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+        embed_dim = embed_dim_ratio * num_joints  #### temporal embed_dim is num_joints * spatial embedding dim ratio
+        out_dim = num_joints * 3  #### output dimension is num_joints * 3
+        ### Temporal patch embedding
+        self.Temporal_pos_embed = nn.Parameter(torch.zeros(1, num_frame, embed_dim))
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer)
+            for i in range(depth)])
+
+        self.Temporal_norm = norm_layer(embed_dim)
+        ####### A easy way to implement weighted mean
+        self.weighted_mean = torch.nn.Conv1d(in_channels=num_frame, out_channels=1, kernel_size=1)
+
+    def forward(self, x):
+        b = x.shape[0]
+        x += self.Temporal_pos_embed
+        x = self.pos_drop(x)
+        for blk in self.blocks:
+            x = blk(x)
+
+        x = self.Temporal_norm(x)
+        ##### x size [b, f, emb_dim], then take weighted mean on frame dimension, we only predict 3D pose of the center frame
+        # x = self.weighted_mean(x)
+        x = x.view(b, opt.frames, -1)
+        return x