Hi, thanks for your wonderful work, I want to ask whether the proposed Micro-Factorized pointwise convolution is a plug and play model? I want to replace pointwise convolution with your Micro-Factorized pointwise convolution in MobileNetV3 .

I cannot find the code of MaxGroupPooling used in line 102 of micronet.py. https://github.com/liyunsheng13/micronet/blob/main/backbone/micronet.py#L102 Can you check?

md5sum * f514ad5f6d691793cfca029f31d87128 micronet-m0.pth f514ad5f6d691793cfca029f31d87128 micronet-m1.pth f514ad5f6d691793cfca029f31d87128 micronet-m2.pth f514ad5f6d691793cfca029f31d87128 micronet-m3.pth All are the m3?

Hi @liyunsheng13,

how to finetune the model. I got the message error:

torch.nn.modules.module.ModuleAttributeError: 'DataParallel' object has no attribute 'classifier'

Thanks

Hi.

I am trying to implement MicroNet in TensorFlow. I am trying to understand the computation differences between SpatialSepConvSF and DepthSpatialSepConv. From the looks of it, their blocks look all the same to me barring the number of output channels that could have been efficiently parameterized through a single class I believe.

The only difference seems to be the ChannelShuffle layer in the former i.e., the SpatialSepConvSF block. ChannelShuffle isn't mentioned in the paper as well.

Could you shed some light?

Hi, I've already read your paper and thought that's a wonderful job. But when I try to read your code, I found those three blocks are not very clear in it. It confuses me cause I want to use them as the part of other models. So it may be rude, but would you please rewrite the code with the 'Micro-Block-A, Micro-Block-B, and Micro-Block-C' description as it is in the paper. Thank you very much!

now the codes of the block looks like this: '''

##StemLayer

A  ## + MFDC + RELU + Shuffle1 + MFPC + act_layer + Shuffle1 + Shuffle2

A  ## + MFDC + act_layer + Shuffle1 + Shuffle2 + MFPC + act_layer +Shuffle1 + Shuffle2

A? ## + MFDC + act_layer + Shuffle1 + Shuffle2 + MFPC + act_layer +Shuffle1 + Shuffle2

B? ## + MFPC + act_layer + Shuffle1 + MFDC + act_layer + Shuffle2 + MFPC + act_layer + Shuffle1 + Shuffle2

C  ## + MFPC + act_layer + Shuffle1 + MFDC + act_layer + Shuffle2 + MFPC + act_layer + Shuffle1 + Shuffle2

C  ## + MFPC + act_layer + Shuffle1 + MFDC + act_layer + Shuffle2 + MFPC + act_layer + Shuffle1 + Shuffle2

C? ## + MFPC + act_layer

'''

but it should be like:

'''

##StemLayer


A  ## + MFDC + RELU + Shuffle1 + MFPC + act_layer + Shuffle1 + Shuffle2

A  ## + MFDC + act_layer + Shuffle1 + Shuffle2 + MFPC + act_layer + Shuffle1 + Shuffle2

B  ## + MFDC + act_layer + Shuffle1 + Shuffle2 + MFPC + act_layer +Shuffle1 + Shuffle2 + MFPC + act_layer 

C  ## + Shuffle1 + MFDC + act_layer + Shuffle2 + MFPC + act_layer + Shuffle1 + Shuffle2 + MFPC + act_layer

C  ## + Shuffle1 + MFDC + act_layer + Shuffle2 + MFPC + act_layer + Shuffle1 + Shuffle2 + MFPC + act_layer

C  ## + Shuffle1 + MFDC + act_layer + Shuffle2 + MFPC + act_layer + Shuffle1 + Shuffle2 + MFPC + act_layer

'''

First of all, thank you for sharing your good research results.

I'm trying to download a pre-trained model, but there seems to be a problem with the address.

I trained about 300 epoches, when I use the model to test , different epoches test results are very different ， sometimes the result is good ， sometimes the result is very bad， how to resolve this problem

Thanks for your great work, but I have trained the train_micronet_m0_4gpu.sh script, but got best Acc@1 45.86 at epoch 596, lower than 46.6 reported in your repo with pytorch 1.7.1 CUDA 10.2 cudnn 7.6.05.

And I did not make any modifications to the training program, is there something wrong with my training program?

hi， thanks so much for your great work. we have some questions about the MicroNet config for Human Pose Estimation task too. so we appreciate it if you could share the code for key point detection. May i ask the specific time for the code to be released?

Hello, I have a question I would like to ask: If the MicroNet network is used for target detection, then as the backbone network, how to modify its output? That is, how to change the one-dimensional output of the classification network to a multi-dimensional output to prepare for the subsequent feature fusion?

My loss down to 0.75 and its hard to converge（150epochs）. I use the M3 model, how can i fix the issue? My data is good and i had done some experiment on some other net.

I am doing related work based on paddleOCR. I replaced the backbone of the CRNN network structure from mobilenet_v3 to micronet. I train on a single machine with 4 cards, the batch_size is 256, and I find that the ips I train are about 60, while the ips of mobilenet_v3 can reach about 200. may I ask if this is normal? And GPU usage has become lower.