In patch merging, suppose an input feature map is reshaped from (B, H×W, C) to (B, H, W, C) and then partitioned into four sub-tensors x0--x3 via slicing (each taking every other row and column).  What is the shape of each sub-tensor? ==> (B, [Fill in the blank], , [Fill in the blank], , [Fill in the blank], ) What is the shape of the concatenated tensor after the line "torch.cat ([x0, x1, x2, x3], -1)"? ==> (B, [Fill in the blank], , [Fill in the blank], , [Fill in the blank], ) 多项填空题