try gtp vit

models.py

@@ -543,6 +543,44 @@ class FouriER(torch.nn.Module):
 
     def forward_tokens(self, x):
         outs = []
+        B, C, H, W = x.shape
+        N = H*W
+        if self.graph_type in ["Semantic", "Mixed"]:
+            # Generate the semantic graph w.r.t. the cosine similarity between tokens
+            # Compute cosine similarity
+            if self.class_token:
+                x_normed = x[:, 1:] / x[:, 1:].norm(dim=-1, keepdim=True)
+            else:
+                x_normed = x / x.norm(dim=-1, keepdim=True)
+            x_cossim = x_normed @ x_normed.transpose(-1, -2)
+            threshold = torch.kthvalue(x_cossim, N-1-self.num_neighbours, dim=-1, keepdim=True)[0] # B,H,1,1 
+            semantic_graph = torch.where(x_cossim>=threshold, 1.0, 0.0)
+            if self.class_token:
+                semantic_graph = semantic_graph - torch.eye(N-1, device=semantic_graph.device).unsqueeze(0)
+            else:
+                semantic_graph = semantic_graph - torch.eye(N, device=semantic_graph.device).unsqueeze(0)
+        
+        if self.graph_type == "None":
+            graph = None
+        else:
+            if self.graph_type == "Spatial":
+                graph = self.spatial_graph.unsqueeze(0).expand(B,-1,-1)#.to(x.device)
+            elif self.graph_type == "Semantic":
+                graph = semantic_graph
+            elif self.graph_type == "Mixed":
+                # Integrate the spatial graph and semantic graph
+                spatial_graph = self.spatial_graph.unsqueeze(0).expand(B,-1,-1).to(x.device)
+                graph = torch.bitwise_or(semantic_graph.int(), spatial_graph.int()).float()
+            
+            # Symmetrically normalize the graph
+            degree = graph.sum(-1) # B, N
+            degree = torch.diag_embed(degree**(-1/2))
+            graph = degree @ graph @ degree
+            
+        if self.token_scale:
+            token_scales = self.token_scales.unsqueeze(0).expand(B,-1).to(x.device)
+        else:
+            token_scales = None
         for idx, block in enumerate(self.network):
             x = block(x)
         # output only the features of last layer for image classification
@@ -868,6 +906,270 @@ def window_reverse(windows, window_size, H, W):
     x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, -1, H, W)
     return x
 
+def propagate(x: torch.Tensor, weight: torch.Tensor, 
+              index_kept: torch.Tensor, index_prop: torch.Tensor, 
+              standard: str = "None", alpha: Optional[float] = 0, 
+              token_scales: Optional[torch.Tensor] = None,
+              cls_token=True):
+    """
+    Propagate tokens based on the selection results.
+    ================================================
+    Args:
+        - x: Tensor([B, N, C]): the feature map of N tokens, including the [CLS] token.
+        
+        - weight: Tensor([B, N-1, N-1]): the weight of each token propagated to the other tokens, 
+                                         excluding the [CLS] token. weight could be a pre-defined 
+                                         graph of the current feature map (by default) or the
+                                         attention map (need to manually modify the Block Module).
+                                         
+        - index_kept: Tensor([B, N-1-num_prop]): the index of kept image tokens in the feature map X
+        
+        - index_prop: Tensor([B, num_prop]): the index of propagated image tokens in the feature map X
+        
+        - standard: str: the method applied to propagate the tokens, including "None", "Mean" and 
+                         "GraphProp"
+        
+        - alpha: float: the coefficient of propagated features
+        
+        - token_scales: Tensor([B, N]): the scale of tokens, including the [CLS] token. token_scales
+                                        is None by default. If it is not None, then token_scales 
+                                        represents the scales of each token and should sum up to N.
+        
+    Return:
+        - x: Tensor([B, N-1-num_prop, C]): the feature map after propagation
+        
+        - weight: Tensor([B, N-1-num_prop, N-1-num_prop]): the graph of feature map after propagation
+        
+        - token_scales: Tensor([B, N-1-num_prop]): the scale of tokens after propagation
+    """
+    
+    B, C, N = x.shape
+    
+    # Step 1: divide tokens
+    if cls_token:
+        x_cls = x[:, 0:1] # B, 1, C
+    x_kept = x.gather(dim=1, index=index_kept.unsqueeze(-1).expand(-1,-1,C)) # B, N-1-num_prop, C
+    x_prop = x.gather(dim=1, index=index_prop.unsqueeze(-1).expand(-1,-1,C)) # B, num_prop, C
+    
+    # Step 2: divide token_scales if it is not None
+    if token_scales is not None:
+        if cls_token:
+            token_scales_cls = token_scales[:, 0:1] # B, 1
+        token_scales_kept = token_scales.gather(dim=1, index=index_kept) # B, N-1-num_prop
+        token_scales_prop = token_scales.gather(dim=1, index=index_prop) # B, num_prop
+    
+    # Step 3: propagate tokens
+    if standard == "None":
+        """
+        No further propagation
+        """
+        pass
+        
+    elif standard == "Mean":
+        """
+        Calculate the mean of all the propagated tokens,
+        and concatenate the result token back to kept tokens.
+        """
+        # naive average
+        x_prop = x_prop.mean(1, keepdim=True) # B, 1, C
+        # Concatenate the average token 
+        x_kept = torch.cat((x_kept, x_prop), dim=1) # B, N-num_prop, C
+            
+    elif standard == "GraphProp":
+        """
+        Propagate all the propagated token to kept token
+        with respect to the weights and token scales.
+        """
+        assert weight is not None, "The graph weight is needed for graph propagation"
+        
+        # Step 3.1: divide propagation weights.
+        if cls_token:
+            index_kept = index_kept - 1 # since weights do not include the [CLS] token
+            index_prop = index_prop - 1 # since weights do not include the [CLS] token
+            weight = weight.gather(dim=1, index=index_kept.unsqueeze(-1).expand(-1,-1,N-1)) # B, N-1-num_prop, N-1
+            weight_prop = weight.gather(dim=2, index=index_prop.unsqueeze(1).expand(-1,weight.shape[1],-1)) # B, N-1-num_prop, num_prop
+            weight = weight.gather(dim=2, index=index_kept.unsqueeze(1).expand(-1,weight.shape[1],-1)) # B, N-1-num_prop, N-1-num_prop
+        else:
+            weight = weight.gather(dim=1, index=index_kept.unsqueeze(-1).expand(-1,-1,N)) # B, N-1-num_prop, N-1
+            weight_prop = weight.gather(dim=2, index=index_prop.unsqueeze(1).expand(-1,weight.shape[1],-1)) # B, N-1-num_prop, num_prop
+            weight = weight.gather(dim=2, index=index_kept.unsqueeze(1).expand(-1,weight.shape[1],-1)) # B, N-1-num_prop, N-1-num_prop
+        
+        # Step 3.2: generate the broadcast message and propagate the message to corresponding kept tokens
+        # Simple implementation
+        x_prop = weight_prop @ x_prop # B, N-1-num_prop, C
+        x_kept = x_kept + alpha * x_prop # B, N-1-num_prop, C
+        
+        """ scatter_reduce implementation for batched inputs
+        # Get the non-zero values
+        non_zero_indices = torch.nonzero(weight_prop, as_tuple=True)
+        non_zero_values = weight_prop[non_zero_indices]
+        
+        # Sparse multiplication
+        batch_indices, row_indices, col_indices = non_zero_indices
+        sparse_matmul = alpha * non_zero_values[:, None] * x_prop[batch_indices, col_indices, :]
+        reduce_indices = batch_indices * x_kept.shape[1] + row_indices
+        
+        x_kept = x_kept.reshape(-1, C).scatter_reduce(dim=0, 
+                                                      index=reduce_indices[:, None], 
+                                                      src=sparse_matmul, 
+                                                      reduce="sum",
+                                                      include_self=True)
+        x_kept = x_kept.reshape(B, -1, C)
+        """
+        
+        # Step 3.3: calculate the scale of each token if token_scales is not None
+        if token_scales is not None:
+            if cls_token:
+                token_scales_cls = token_scales[:, 0:1] # B, 1
+                token_scales = token_scales[:, 1:]
+            token_scales_kept = token_scales.gather(dim=1, index=index_kept) # B, N-1-num_prop
+            token_scales_prop = token_scales.gather(dim=1, index=index_prop) # B, num_prop
+            token_scales_prop = weight_prop @ token_scales_prop.unsqueeze(-1) # B, N-1-num_prop, 1
+            token_scales = token_scales_kept + alpha * token_scales_prop.squeeze(-1) # B, N-1-num_prop
+            if cls_token:
+                token_scales = torch.cat((token_scales_cls, token_scales), dim=1) # B, N-num_prop
+    else:
+        assert False, "Propagation method \'%f\' has not been supported yet." % standard
+    
+    
+    if cls_token:
+        # Step 4： concatenate the [CLS] token and generate returned value
+        x = torch.cat((x_cls, x_kept), dim=1) # B, N-num_prop, C
+    else:
+        x = x_kept
+    return x, weight, token_scales
+
+
+
+def select(weight: torch.Tensor, standard: str = "None", num_prop: int = 0, cls_token = True):
+    """
+    Select image tokens to be propagated. The [CLS] token will be ignored. 
+    ======================================================================
+    Args:
+        - weight: Tensor([B, H, N, N]): used for selecting the kept tokens. Only support the
+                                        attention map of tokens at the moment.
+        
+        - standard: str: the method applied to select the tokens
+        
+        - num_prop: int: the number of tokens to be propagated
+        
+    Return:
+        - index_kept: Tensor([B, N-1-num_prop]): the index of kept tokens 
+        
+        - index_prop: Tensor([B, num_prop]): the index of propagated tokens
+    """
+    
+    assert len(weight.shape) == 4, "Selection methods on tensors other than the attention map haven't been supported yet."
+    B, H, N1, N2 = weight.shape
+    assert N1 == N2, "Selection methods on tensors other than the attention map haven't been supported yet."
+    N = N1
+    assert num_prop >= 0, "The number of propagated/pruned tokens must be non-negative."
+    
+    if cls_token:
+        if standard == "CLSAttnMean":
+            token_rank = weight[:,:,0,1:].mean(1)
+            
+        elif standard == "CLSAttnMax":
+            token_rank = weight[:,:,0,1:].max(1)[0]
+                
+        elif standard == "IMGAttnMean":
+            token_rank = weight[:,:,:,1:].sum(-2).mean(1)
+        
+        elif standard == "IMGAttnMax":
+            token_rank = weight[:,:,:,1:].sum(-2).max(1)[0]
+                
+        elif standard == "DiagAttnMean":
+            token_rank = torch.diagonal(weight, dim1=-2, dim2=-1)[:,:,1:].mean(1)
+            
+        elif standard == "DiagAttnMax":
+            token_rank = torch.diagonal(weight, dim1=-2, dim2=-1)[:,:,1:].max(1)[0]
+            
+        elif standard == "MixedAttnMean":
+            token_rank_1 = torch.diagonal(weight, dim1=-2, dim2=-1)[:,:,1:].mean(1)
+            token_rank_2 = weight[:,:,:,1:].sum(-2).mean(1)
+            token_rank = token_rank_1 * token_rank_2
+            
+        elif standard == "MixedAttnMax":
+            token_rank_1 = torch.diagonal(weight, dim1=-2, dim2=-1)[:,:,1:].max(1)[0]
+            token_rank_2 = weight[:,:,:,1:].sum(-2).max(1)[0]
+            token_rank = token_rank_1 * token_rank_2
+            
+        elif standard == "SumAttnMax":
+            token_rank_1 = torch.diagonal(weight, dim1=-2, dim2=-1)[:,:,1:].max(1)[0]
+            token_rank_2 = weight[:,:,:,1:].sum(-2).max(1)[0]
+            token_rank = token_rank_1 + token_rank_2
+            
+        elif standard == "CosSimMean":
+            weight = weight[:,:,1:,:].mean(1)
+            weight = weight / weight.norm(dim=-1, keepdim=True)
+            token_rank = -(weight @ weight.transpose(-1, -2)).sum(-1)
+        
+        elif standard == "CosSimMax":
+            weight = weight[:,:,1:,:].max(1)[0]
+            weight = weight / weight.norm(dim=-1, keepdim=True)
+            token_rank = -(weight @ weight.transpose(-1, -2)).sum(-1)
+            
+        elif standard == "Random":
+            token_rank = torch.randn((B, N-1), device=weight.device)
+                
+        else:
+            print("Type\'", standard, "\' selection not supported.")
+            assert False
+        
+        token_rank = torch.argsort(token_rank, dim=1, descending=True) # B, N-1
+        index_kept = token_rank[:, :-num_prop]+1 # B, N-1-num_prop
+        index_prop = token_rank[:, -num_prop:]+1 # B, num_prop
+            
+    else:
+        if standard == "IMGAttnMean":
+            token_rank = weight.sum(-2).mean(1)
+        
+        elif standard == "IMGAttnMax":
+            token_rank = weight.sum(-2).max(1)[0]
+                
+        elif standard == "DiagAttnMean":
+            token_rank = torch.diagonal(weight, dim1=-2, dim2=-1).mean(1)
+            
+        elif standard == "DiagAttnMax":
+            token_rank = torch.diagonal(weight, dim1=-2, dim2=-1).max(1)[0]
+            
+        elif standard == "MixedAttnMean":
+            token_rank_1 = torch.diagonal(weight, dim1=-2, dim2=-1).mean(1)
+            token_rank_2 = weight.sum(-2).mean(1)
+            token_rank = token_rank_1 * token_rank_2
+            
+        elif standard == "MixedAttnMax":
+            token_rank_1 = torch.diagonal(weight, dim1=-2, dim2=-1).max(1)[0]
+            token_rank_2 = weight.sum(-2).max(1)[0]
+            token_rank = token_rank_1 * token_rank_2
+            
+        elif standard == "SumAttnMax":
+            token_rank_1 = torch.diagonal(weight, dim1=-2, dim2=-1).max(1)[0]
+            token_rank_2 = weight.sum(-2).max(1)[0]
+            token_rank = token_rank_1 + token_rank_2
+            
+        elif standard == "CosSimMean":
+            weight = weight.mean(1)
+            weight = weight / weight.norm(dim=-1, keepdim=True)
+            token_rank = -(weight @ weight.transpose(-1, -2)).sum(-1)
+        
+        elif standard == "CosSimMax":
+            weight = weight.max(1)[0]
+            weight = weight / weight.norm(dim=-1, keepdim=True)
+            token_rank = -(weight @ weight.transpose(-1, -2)).sum(-1)
+            
+        elif standard == "Random":
+            token_rank = torch.randn((B, N-1), device=weight.device)
+                
+        else:
+            print("Type\'", standard, "\' selection not supported.")
+            assert False
+        
+        token_rank = torch.argsort(token_rank, dim=1, descending=True) # B, N-1
+        index_kept = token_rank[:, :-num_prop] # B, N-1-num_prop
+        index_prop = token_rank[:, -num_prop:] # B, num_prop
+    return index_kept, index_prop
+
 class PoolFormerBlock(nn.Module):
     """
     Implementation of one PoolFormer block.
@@ -910,13 +1212,17 @@ class PoolFormerBlock(nn.Module):
             self.layer_scale_2 = nn.Parameter(
                 layer_scale_init_value * torch.ones((dim)), requires_grad=True)
 
-    def forward(self, x):
+    def forward(self, x, graph):
         B, C, H, W = x.shape
         x_windows = window_partition(x, self.window_size)
         x_windows = x_windows.view(-1, self.window_size * self.window_size, C)
         attn_windows = self.token_mixer(x_windows, mask=self.attn_mask)
         attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
         x_attn = window_reverse(attn_windows, self.window_size, H, W)
+        index_kept, index_prop = select(x_attn, standard="MixedAttnMax", num_prop=0,
+                                            cls_token=False)
+        x, weight, token_scales = propagate(x, weight, index_kept, index_prop, standard="GraphProp",
+                                        alpha=0.1, token_scales=token_scales, cls_token=False)
         if self.use_layer_scale:
             x = x + self.drop_path(
                 self.layer_scale_1.unsqueeze(-1).unsqueeze(-1)