flops的计算问题

[223hzz]

Hi author, I really like your code, recently I was using it and I realized that the fvcore function is not provided in your code when calculating flops, e.g. : DilateAttention class “q @ k” or “attn @ v “ for multiplication or “nn.unflod” for flops.
” k = self.unfold(k)
attn = (q @ k) * self.scale # B,h,N,1,kk
v = self.unfold(v) # B,h,N,kk,d
x = (attn @ v)“
”
attn = (q @ k.transpose(-2, -1))
x = (attn @ v).transpose(1, 2)
“
So are the flops calculated this way accurate?

[JIAOJIAYUASD]

Hello, thank you for your interest in my code. I haven't updated this code repository for a long time.

In fact, during the design process of the model, I did not pay attention to how fvcore calculated the parameters and flops of the model. I followed other people's code. UniFormer.

nn.fold does not produce calculations or parameters, but only increases memory latency.