acheun9/pytorch-implementation-of-mobile-former

Simple implementation of Mobile-Former on Pytorch

pytorch-implementation-of-mobile-former's Issues

some question

class Mobile(nn.Module):
def init(self, ks, inp, hid, out, se, stride, dim, reduction=4, k=2):

hi, call you tell me, the k value why equal to 2, What is it used for

two issues

Great job!
However, I think there are probably two tiny issues in you code.

The first one is in bridge.py(line 24 & line 53). I think there are some differences in the following two lines of code

x = x.reshape(b, c, h*w).transpose(1,2).unsqueeze(1)
x = x.contiguous().view(b, h * w, c).unsqueeze(1)

May be the first line is correct?

The second one is in config.py.Accroding to the original paper, in page 13,

Figure 7. Visualization of cross attention on the two-way bridge: Mobile→Former and Mobile←Former. Mobile-Former-294M is used,which includes 6 tokens (each corresponds to a column) and 11 Mobile-Former blocks (block 2–12) across 4 stages. Each block has two attention heads that are visualized in two rows. Attention in Mobile→Former (left half) is normalized over pixels, showing the focused region per token. Attention in Mobile←Former (right half) is normalized over tokens showing the contribution per token at each pixel.

But in config.py, there are some stages with only one head.

I'm not sure whether the above is correct. Looking forward to your reply!

Preweights?

Hi, I want to extend the model on my own task, will you release pre-trained weights?

Model

您好，问一下，这个实现中关于transformer的部分是没有考虑位置编码吗？

配置

请问作者是什么电脑配置跑这个模型的啊，2080ti能带动吗？

Not converge

Great job！
However, i try the training on imagent and it does not converge.
I also try another implement https://github.com/slwang9353/MobileFormer and it does not converge either.
Does anyone successfully reproduce the results in the paper?