最后一个epoch验证不完，以及shape问题

[sdreamforchen]

问题确认 Search before asking

• 我已经查询历史issue，没有发现相似的bug。I have searched the issues and found no similar bug report.

Bug组件 Bug Component

No response

Bug描述 Describe the Bug

您好
1 在训练RT-DETR的时候，老是最后的eval，只eval到1000就不动了，每次都需要手动eval 最后的权重，有点麻烦。
2 在rtdetr_transformer.py里面使用x=self.lamda * F.softmax(x) + (1 - self.lamda) * F.sigmoid(x)，在169行和170行的attention_weights = F.sigmoid(attention_weights).reshape([bs, Len_q, self.num_heads, self.num_levels, self.num_points])会报 shape错误。完整代码是
class Act(nn.Layer):
def init(self, step=30000, r0=1.0):
super(Act, self).init()
self.iter = step+1
self.total_step=step
self.r0 = r0

    self.lamda = paddle.full([1],1,"float32")
    
def forward(self, x):
    
    if self.training:
        
        if self.iter > 0:
            self.iter=self.iter - 1
        #    a = self.iter
        
        
        #print(self.iter)
        
        x1 = F.softmax(x)
        x2 = F.sigmoid(x)
                                 
        if self.iter<=self.total_step:
            self.lamda = paddle.to_tensor(self.r0*self.iter/self.total_step)
        else:
            self.lamda = paddle.to_tensor(self.r0)
        x = paddle.to_tensor(self.lamda * x1 + (1 - self.lamda) * x2)
    else:
        #x = self.norm2(x)
        if self.iter==0:
            x = paddle.to_tensor(F.sigmoid(x))
        else:
            x = paddle.to_tensor(self.lamda * F.softmax(x) + (1 - self.lamda) * F.sigmoid(x))
    #mm=self.lamda
    return x

class PPMSDeformableAttention(MSDeformableAttention):
........
attention_weights = Act(attention_weights).reshape([bs, Len_q, self.num_heads, self.num_levels, self.num_points])
.......

复现环境 Environment

最新版

Bug描述确认 Bug description confirmation

• 我确认已经提供了Bug复现步骤、代码改动说明、以及环境信息，确认问题是可以复现的。I confirm that the bug replication steps, code change instructions, and environment information have been provided, and the problem can be reproduced.

是否愿意提交PR？ Are you willing to submit a PR?

• 我愿意提交PR！I'd like to help by submitting a PR!