mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

yolov3-spp？ #29

New Issue
Closed
opened 2020-06-04 03:10:52 +02:00 by ou525 · 14 comments
ou525 commented 2020-06-04 03:10:52 +02:00 (Migrated from github.com)
Copy Link
Copy Source

Thank you very much for your work.
Have you tested the effect of yolov3-spp network deployment on tkDnn, is it consistent with darknet?

Thank you very much for your work. Have you tested the effect of yolov3-spp network deployment on tkDnn, is it consistent with darknet?
ceccocats commented 2020-06-04 12:48:34 +02:00 (Migrated from github.com)
Copy Link
Copy Source

If yolov3-spp use the same layers and parameters of yolo3 it is supported by our Darknet cfg parser

If yolov3-spp use the same layers and parameters of yolo3 it is supported by our Darknet cfg parser
ou525 commented 2020-06-05 05:34:52 +02:00 (Migrated from github.com)
Copy Link
Copy Source

I have converted my yolov3-spp network successfully (only modify the category and my own dataset), and tested some pictures, and found that some output is not consistent.
I don't know where is the problem

I have converted my yolov3-spp network successfully (only modify the category and my own dataset), and tested some pictures, and found that some output is not consistent. I don't know where is the problem
ceccocats commented 2020-06-05 10:56:25 +02:00 (Migrated from github.com)
Copy Link
Copy Source

Can you list all the steps you have done?
You should:

  • get your net.cfg and net.weights in darknet
  • export the weights using our darknet fork
  • make a test in tkdnn or modify an existing one that take your net.cfg and your exported weights.
  • execute the test and check if it is correct
  • use the generated rt file in demo or whatsoever
Can you list all the steps you have done? You should: - get your net.cfg and net.weights in darknet - export the weights using our darknet fork - make a test in tkdnn or modify an existing one that take your net.cfg and your exported weights. - execute the test and check if it is correct - use the generated rt file in demo or whatsoever
ou525 commented 2020-06-05 11:05:47 +02:00 (Migrated from github.com)
Copy Link
Copy Source

Yes, I have completed all the steps, otherwise I cannot continue.
new a test cpp
#include
#include
#include "tkdnn.h"
#include "test.h"
#include "DarknetParser.h"

int main() {
std::string bin_path = "yolo_spp";
std::vectorstd::string input_bins = {
bin_path + "/layers/input.bin"
};
std::vectorstd::string output_bins = {
bin_path + "/debug/layer89_out.bin",
bin_path + "/debug/layer101_out.bin",
bin_path + "/debug/layer113_out.bin"
};
std::string wgs_path = bin_path + "/layers";
std::string cfg_path = "my.cfg";
std::string name_path = "my.txt";
downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/d97CFzYqCPCp5Hg/download");

// parse darknet network
tk::dnn::Network *net = tk::dnn::darknetParser(cfg_path, wgs_path, name_path);
net->print();

//convert network to tensorRT
tk::dnn::NetworkRT *netRT = new tk::dnn::NetworkRT(net, net->getNetworkRTName(bin_path.c_str()));

int ret = testInference(input_bins, output_bins, net, netRT);
net->releaseLayers();
delete net;
delete netRT;
return ret;

}

Yes, I have completed all the steps, otherwise I cannot continue. new a test cpp #include<iostream> #include<vector> #include "tkdnn.h" #include "test.h" #include "DarknetParser.h" int main() { std::string bin_path = "yolo_spp"; std::vector<std::string> input_bins = { bin_path + "/layers/input.bin" }; std::vector<std::string> output_bins = { bin_path + "/debug/layer89_out.bin", bin_path + "/debug/layer101_out.bin", bin_path + "/debug/layer113_out.bin" }; std::string wgs_path = bin_path + "/layers"; std::string cfg_path = "my.cfg"; std::string name_path = "my.txt"; downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/d97CFzYqCPCp5Hg/download"); // parse darknet network tk::dnn::Network *net = tk::dnn::darknetParser(cfg_path, wgs_path, name_path); net->print(); //convert network to tensorRT tk::dnn::NetworkRT *netRT = new tk::dnn::NetworkRT(net, net->getNetworkRTName(bin_path.c_str())); int ret = testInference(input_bins, output_bins, net, netRT); net->releaseLayers(); delete net; delete netRT; return ret; }
ceccocats commented 2020-06-05 16:54:28 +02:00 (Migrated from github.com)
Copy Link
Copy Source

Are you sure the output layers are the same? If you add or remove layers the output number will change.
To be sure you can set this as output-bins:

std::vectorstd::string output_bins = {
bin_path + "/debug/layer89_out.bin",
bin_path + "/debug/layer101_out.bin",
bin_path + "/layers/output.bin"
};

Output.bin is the last layer output, so at least the last yolo layer should be right.
For the others yolos you must find the correct number when you are exporting from darknet

Are you sure the output layers are the same? If you add or remove layers the output number will change. To be sure you can set this as output-bins: ``` std::vectorstd::string output_bins = { bin_path + "/debug/layer89_out.bin", bin_path + "/debug/layer101_out.bin", bin_path + "/layers/output.bin" }; ``` Output.bin is the last layer output, so at least the last yolo layer should be right. For the others yolos you must find the correct number when you are exporting from darknet
ou525 commented 2020-06-06 09:26:35 +02:00 (Migrated from github.com)
Copy Link
Copy Source

The output layer should be correct, I did not modify the network structure. Does the nms implementation affect the results

The output layer should be correct, I did not modify the network structure. Does the nms implementation affect the results
ceccocats commented 2020-06-06 11:13:01 +02:00 (Migrated from github.com)
Copy Link
Copy Source

The nms affect the results only if you do the postprocessing that is not executed in the test.
Could you paste the cfg? Maybe also weights.
When I have this problems i remove layers until I find what is the layer that create the error.
Maybe your cfg contains an option not parsed.

The nms affect the results only if you do the postprocessing that is not executed in the test. Could you paste the cfg? Maybe also weights. When I have this problems i remove layers until I find what is the layer that create the error. Maybe your cfg contains an option not parsed.
ou525 commented 2020-06-08 04:08:42 +02:00 (Migrated from github.com)
Copy Link
Copy Source

thank you，i uploaded the cfg and weights files, you can use it as a reference test.
https://drive.google.com/file/d/10vGFfsRjCrP-jKU7L5vrkjZtBF-gV5LN/view?usp=sharing
https://drive.google.com/file/d/1eUq03hKS_po7HDk6X67QSFr10t3xYizf/view?usp=sharing

thank you，i uploaded the cfg and weights files, you can use it as a reference test. https://drive.google.com/file/d/10vGFfsRjCrP-jKU7L5vrkjZtBF-gV5LN/view?usp=sharing https://drive.google.com/file/d/1eUq03hKS_po7HDk6X67QSFr10t3xYizf/view?usp=sharing
NickiBD commented 2020-06-09 00:04:48 +02:00 (Migrated from github.com)
Copy Link
Copy Source

@ceccocats ,Hi,
Thanks lot for your great work .I had a question regarding execution of the changed .cpp for parser .I have created a new cpp file (parser ) and changed the out puts for the custom cfg in darknet (yolov3-spp) and also exported the weights (bin files ). However, how should I execute the changed .cpp file to generate the rt file for the demo. I used g++ ..... to compile the .cpp file but it gave me include errors (cuda.h , tkdnn no such file or directory ) .I would be really grateful if you could help me with this . Thanks in advance .

@ceccocats ,Hi, Thanks lot for your great work .I had a question regarding execution of the changed .cpp for parser .I have created a new cpp file (parser ) and changed the out puts for the custom cfg in darknet (yolov3-spp) and also exported the weights (bin files ). However, how should I execute the changed .cpp file to generate the rt file for the demo. I used g++ ..... to compile the .cpp file but it gave me include errors (cuda.h , tkdnn no such file or directory ) .I would be really grateful if you could help me with this . Thanks in advance .
ceccocats commented 2020-06-12 14:32:12 +02:00 (Migrated from github.com)
Copy Link
Copy Source

thank you，i uploaded the cfg and weights files, you can use it as a reference test.
https://drive.google.com/file/d/10vGFfsRjCrP-jKU7L5vrkjZtBF-gV5LN/view?usp=sharing
https://drive.google.com/file/d/1eUq03hKS_po7HDk6X67QSFr10t3xYizf/view?usp=sharing

I still can't access the files

> thank you，i uploaded the cfg and weights files, you can use it as a reference test. > https://drive.google.com/file/d/10vGFfsRjCrP-jKU7L5vrkjZtBF-gV5LN/view?usp=sharing > https://drive.google.com/file/d/1eUq03hKS_po7HDk6X67QSFr10t3xYizf/view?usp=sharing I still can't access the files
ou525 commented 2020-06-13 05:06:00 +02:00 (Migrated from github.com)
Copy Link
Copy Source

sorry，please try again
https://drive.google.com/file/d/1eUq03hKS_po7HDk6X67QSFr10t3xYizf/view?usp=sharing
https://drive.google.com/file/d/10vGFfsRjCrP-jKU7L5vrkjZtBF-gV5LN/view?usp=sharing

sorry，please try again https://drive.google.com/file/d/1eUq03hKS_po7HDk6X67QSFr10t3xYizf/view?usp=sharing https://drive.google.com/file/d/10vGFfsRjCrP-jKU7L5vrkjZtBF-gV5LN/view?usp=sharing
ceccocats commented 2020-06-14 13:11:10 +02:00 (Migrated from github.com)
Copy Link
Copy Source

I executed it without any problems, but pull the repo, some tensorRT issue can be due to different include and library dirs, the last commits solves this cmake issue.
I changed the output bins that in your network are different, the simplier way to check what de bug layers are the YOLO ones is to check what layers weights starts with "g":

build$ ls yolov3-sppx/layers/g*
yolov3-sppx/layers/g101.bin  yolov3-sppx/layers/g113.bin  yolov3-sppx/layers/g89.bin

This is the code:

#include<iostream>
#include<vector>
#include "tkdnn.h"
#include "test.h"
#include "DarknetParser.h"

int main() {
    std::string bin_path  = "yolov3-sppx";
    std::vector<std::string> input_bins = { 
        bin_path + "/layers/input.bin"
    };
    std::vector<std::string> output_bins = {
        bin_path + "/debug/layer89_out.bin",
        bin_path + "/debug/layer101_out.bin",
        bin_path + "/debug/layer113_out.bin"
    };
    std::string wgs_path  = bin_path + "/layers";
    std::string cfg_path  = "../tests/darknet/cfg/yolov3-sppx.cfg";
    std::string name_path = "../tests/darknet/names/coco4.names";
    //downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/jPXmHyptpLoNdNR/download");

    // parse darknet network
    tk::dnn::Network *net = tk::dnn::darknetParser(cfg_path, wgs_path, name_path);
    net->print();

    //convert network to tensorRT
    tk::dnn::NetworkRT *netRT = new tk::dnn::NetworkRT(net, net->getNetworkRTName(bin_path.c_str()));
    
    int ret = testInference(input_bins, output_bins, net, netRT);    
    net->releaseLayers();
    delete net;
    delete netRT;
    return ret;
}
I executed it without any problems, but pull the repo, some tensorRT issue can be due to different include and library dirs, the last commits solves this cmake issue. I changed the output bins that in your network are different, the simplier way to check what de bug layers are the YOLO ones is to check what layers weights starts with "g": ``` build$ ls yolov3-sppx/layers/g* yolov3-sppx/layers/g101.bin yolov3-sppx/layers/g113.bin yolov3-sppx/layers/g89.bin ``` This is the code: ``` #include<iostream> #include<vector> #include "tkdnn.h" #include "test.h" #include "DarknetParser.h" int main() { std::string bin_path = "yolov3-sppx"; std::vector<std::string> input_bins = { bin_path + "/layers/input.bin" }; std::vector<std::string> output_bins = { bin_path + "/debug/layer89_out.bin", bin_path + "/debug/layer101_out.bin", bin_path + "/debug/layer113_out.bin" }; std::string wgs_path = bin_path + "/layers"; std::string cfg_path = "../tests/darknet/cfg/yolov3-sppx.cfg"; std::string name_path = "../tests/darknet/names/coco4.names"; //downloadWeightsifDoNotExist(input_bins[0], bin_path, "https://cloud.hipert.unimore.it/s/jPXmHyptpLoNdNR/download"); // parse darknet network tk::dnn::Network *net = tk::dnn::darknetParser(cfg_path, wgs_path, name_path); net->print(); //convert network to tensorRT tk::dnn::NetworkRT *netRT = new tk::dnn::NetworkRT(net, net->getNetworkRTName(bin_path.c_str())); int ret = testInference(input_bins, output_bins, net, netRT); net->releaseLayers(); delete net; delete netRT; return ret; } ```
ou525 commented 2020-06-15 03:51:32 +02:00 (Migrated from github.com)
Copy Link
Copy Source

@ceccocats sorry, this code seems to be basically the same as mine

@ceccocats sorry, this code seems to be basically the same as mine
mive93 commented 2020-09-11 09:21:37 +02:00 (Migrated from github.com)
Copy Link
Copy Source

Closing for now,
feel free to reopen.

Closing for now, feel free to reopen.
This repo is archived. You cannot comment on issues.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
bug
duplicate
enhancement
explanation
help wanted
invalid
question
wontfix
No Label
Milestone
No items
No Milestone
Assignees
Clear assignees
eeeck (Erick)
No Assignees
1 Participants
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: mmr/tkDNN#29
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?