shoutoutyangjie / mobileone Goto Github PK

o = deploy_model(x)
this "x" should't have.

hi，when i read your code，i find that a question about merge conv and bn.
when merge conv and bn：

why your code ：
std = (running_var + eps).sqrt()
t = (gamma / std).reshape(-1, 1, 1, 1)
return kernel * t, beta - running_mean * gamma / std

why you not use conv's bias, i think the result should be:
weight = conv_weight * bn_weight.view(out_channels, 1, 1, 1) / bn_std.view(out_channels, 1, 1, 1)

     bias = bn_weight * (conv_bias - bn_mean) / bn_std + bn_bias

Hi, I find that using repvgg_model_convert function, the output of the model and the deployed model are very different using the same input tensor. How did you check the model and depolyment model results consistency ?

self.stage0 look like normal stem, rather than mobileOne block in paper. Any reason why?

By the way, use mobileOne block as stem might cause huge acc drop in cifar100 (50%+)

Hi @shoutOutYangJie, I've noticed that you have a bug in your MobileOne implementation:

You are adding DW BN layer based on in_channels == out_channels condition but it should be always added because for DW part input channels are always equal to output channels. This condition should be only checked for PW part as there might be channel change.

I have no idea about the parameters of mobileone_s2. Could you help offer the parameters to me?

I like your work very much, it gave me a lot of inspiration, because I am a beginner, some codes do not understand, I would like to ask, in this part of the code, I commented "model = copy.deepcopy(model)" Does this line of code affect the final result? Why can't reparameterization be done directly

Hello,I wanna train my model,but I don't know the format of the dataset.
Could you please provide it?Thank you.
The project brought me a lot of help,I really need it.
Please!

Hi,
Thank you very much for the implementation.
I would like to know if you still have the graphs of the full model validation loss and the inference (ready to deploy) model validation loss.
I would like to know how both models behaved in the very early stage of training (first 20 epochs).
Thank you very much for considering my request :)

大佬有在iphone上实测推理速度吗？

Could you share the weights you trained on BaiduyunDisk or change the access permission of the file shared on google disk?