mobilenetv2 + detection training problems

[jacky4323]

thanks for your great work!!

I have some questions. I use this model(multiplier = 1.0) to train my detection model and I have to resize my input probably to 300x300 or 512x512 or 416x416,but the pretrained model you provided is 224x224,If this will cause some problems to train detection model ?

thanks for your suggestions.

[liangfu]

I have trained a ssd model with the features. Even the input image size is as large as 512x1024, the pretrained model works fine.

[jacky4323]

ok thanks for suggestions, I will try .

I have trained mobilenet SSD in the past, Shall I fix the parameters in mobilenetv2 and finetune ssd network? or some important things I should know? many thanks.

[liangfu]

i my observation, there is no need to fix the parameters in mobilenetv2 while training ssd.

[jacky4323]

Hi , @liangfu

I have train on pascol voc dataset , it just can achieve 68 mAP , what is the limit of this model?(75%、80%?) thanks!!

[liangfu]

can you provide more detail on how did u train your model? and what exactly do you mean by referring to the limit of the model?

[jacky4323]

I am still training now , I'm not very rigorous to tune the parameters, so after training complete I will back :) I also notice that mxnet gluon model zoo have some pretrained model,Could I download the params file and use your symbol to train? or I still have to rename the layers' name due to different name for pretrained model

thank you for your patience :)

[liangfu]

i have not tried the pretrained gluon model, but based on my previous experience with gluon models, i think there is a need to rename the model names, due to incompatibility.

why don't you download the models in the models folder to fine tune your detection network?

[jacky4323]

I have downloaded the models in the models folder. in the begining,the performance isn't good at PacolVoc,now I strictly tune the parameters and still training the model now so I don't know the performance. gluon model zoo also has multiplier = 0.5 0.75 I am interested(due to less computation) thanks!!

[liangfu]

that make sense. i can upload pretrained model with multiplier with 0.75 and 0.5 then, so that more people could benefit from this.

[jacky4323]

Hi @liangfu

thanks for your kindness, I think it will be very helpful !!

[jacky4323]

@liangfu Hi,

I'm back :) I use https://github.com/zhreshold/mxnet-ssd to train mobilenetv2 model but in input data shape 300x300 and initialize learning rate 0.002 and scale 0.1 it at epoch 80 and 160 the experiment result at PascolVoc just can achieve 60 mAP , is that resonable? or Could this model achieve 75 mAP?

Thanks!

[liangfu]

I think mobilenetv2 based ssdlite should be reproducible if you stick on the details described in the paper.

Can you tell which layers did you feed into the detection layers? I think they supposed to be the element-wise added shortcut layers.

[jacky4323]

Hi,

seq-5-block0-exp-batchnorm、last-1x1-conv-batchnorm and four extra layers(all these layers is depthwise followed by 1 × 1 projection) to feed into detection layers(prediction confidence and location use normal convolution)

What is the meaning of reproducible ? I only train on PascolVoc and I didn't knowwhat performance it can achieve(70 mAP or 75 mAP or 80 mAP)

thank you!!

[liangfu]

I mean to reproduce the mAP results stated in the paper.

IMHO, "the last layer" you highlighted means the shortcut layers in mobilenetv2, not the batchnorm layer.