Changes to made for more than 5 classes

[PiyalGeorge]

Thanks @Ghustwb for this repo. Finally it worked. I trained a model in with 5 classes and it worked. But how to make it work for 7 classes? I mean where all we have to make changes?

[Ghustwb]

1、pluginImplement.cpp reshape layer 2、pluginImplement.h softmax 3、pluginImplement.cpp detection_out layer you could replace "5" with "7"

[PiyalGeorge]

@Ghustwb , extremely Sorry for the late reply. I tried changing following things, but still getting error:- In pluginImplement.cpp line number 369 - mMbox_conf_reshape = std::unique_ptr<Reshape<5>>(new Reshape<5>()); line number 392 - params.numClasses = 5; line number 652 - mMbox_conf_reshape = std::unique_ptr<Reshape<5>>(new Reshape<5>(serialData, serialLength)); In pluginImplement.h line number 150 - assert((inputs[0].d[0])*(inputs[0].d[1]) % 5 == 0); line number 320 - std::unique_ptr<Reshape<5>> mMbox_conf_reshape{ nullptr };

then i clean, cmake, make. when i run the command './build/bin/mobileNet' , i get following error-

attempting to open cache file ../../model/MobileNetSSD_deploy.caffemodel.1.tensorcache
cache file not found, profiling network model
../../model/MobileNetSSD_deploy_iplugin.prototxt
../../model/MobileNetSSD_deploy.caffemodel
CaffeParser: Could not open file ../../model/MobileNetSSD_deploy.caffemodel
CaffeParser: Could not parse model file
mobileNet: /home/nvidia/MobileNet-SSD-TensorRT-master/tensorNet.cpp:105: bool TensorNet::caffeToTRTModel(const char*, const char*, const std::vector<std::__cxx11::basic_string<char> >&, unsigned int, std::ostream&): Assertion `blobNameToTensor != nullptr' failed.
Aborted

I also saw this line in pluginImplement.h line number 167, 168 - //cudaSoftmax( 8732 21, 21, (float ) inputs, static_cast<float >(outputs)); cudaSoftmax( 1917 5, 5, (float ) inputs, static_cast<float >(outputs));

Do i need to make any change in above line 167, 168 in pluginImplement.h? what is 8732 and 1917? so i need to change that? Is there something else i need to change? Please help

[PiyalGeorge]

@Ghustwb Please help

[Ghustwb]

Sorry for the late reply

//cudaSoftmax( 8732 *21, 21, (float *) *inputs, static_cast<float *>(*outputs));
cudaSoftmax( 1917 *5, 5, (float *) *inputs, static_cast<float *>(*outputs));

In SSD，the backbone is VGG16，Conv4_3 (38*38)can get 4 boxes of different widths and heights，Conv7 (19*19) get 6 boxes，Conv8_2 and Conv9_2 get 6 boxes，Conv 10_2 and Conv11_2 get 4 boxes. So,

38*38*4+19*19*6+10*10*6+5*5*6+3*3*4+4 = 8732。

In MobileNet-SSD,the backbone is mobileNet,so the number of boxes is different with SSD. you can read the paper of mobileNet.

If you use 2 classes, you should chage it

cudaSoftmax( 1917 *2, 2, (float *) *inputs, static_cast<float *>(*outputs));

[PiyalGeorge]

Thanks @Ghustwb , Finally figured out. Actually it was simple, but took a lot of time to figure it out. For guys who are still looking for the same answer, give me a smiley here, :smiley: , Just kidding.

For changing classes from 5 to our custom number of classes make changes in following files:

(custom number of classes mentioned here is including background)

In pluginImplement.cpp: make change in following lines from 5 to your custom number of classes: 369 392 652

In pluginImplement.h: make change in following lines from 5 to your custom number of classes: 150 168 320

In MobileNetSSD_deploy_iplugin.prototxt: Change lines, where n is number of classes including background

650 - 15 to n 3 701 - 30 to n 6 752 - 30 to n 6 803 - 30 to n 6 854 - 30 to n 6 905 - 30 to n 6

example, below is for changing 5(including background) classes to 3(including background)

650 - 15 to 9 701 - 30 to 18 752 - 30 to 18 803 - 30 to 18 854 - 30 to 18 905 - 30 to 18

Thanks

[vagrant-drift]

@PiyalGeorge just make changes in these files? Need to retrain?

[PiyalGeorge]

Hi @xiongdahua555 , the caffemodel(Mobilenet-ssd) in this repo by Ghustwb is having 5 classes. if you have a custom class trained with similar number of classes, just replace Ghustwb's model with that. and do make changes in above files. Also i got an error when tried the same with 21 classes model(Maybe more changes might be needed, but i didn't checked into that since i already got results with 3). When i tried with 3 classes the above changes gave successful results. Thank you