search

ceccocats / tkDNN

Deep neural network library and toolkit to do high performace inference on NVIDIA jetson platforms
GNU General Public License v2.0
718 stars 209 forks source link

After generating .rt, there is no check boxes when using demo.exe to test #256

Closed B1ackg0d closed 3 years ago

B1ackg0d commented 3 years ago

After the .rt is generated, the display is as follows： image

mive93 commented 3 years ago

It seems like the debug file are missing. Does the demo work? Which steps did you follow to reach that result?

B1ackg0d commented 3 years ago

Hi,mive93!I thought the debug file was not necessary, so I didn’t care about it. I first used darknet to export the weights. Then I changed the names and cfg files under tkdnn/test/darknet. Then use test_yolo4.exe to generate the rt file, it is as shown in the figure above. Then I run demo.exe to test rt file, and I find that there is no check boxes.

mive93 commented 3 years ago

You should export the debug and check if the conversion works fine.

B1ackg0d commented 3 years ago

This is what it says in the document，‘if you need debug export of every layer ouput you MUST compile in CPU mode’,I’m not sure how to compile with cpu, so I didn’t use.

mive93 commented 3 years ago

It's quite easy, just leave GPU=0 in darknet makefile

B1ackg0d commented 3 years ago

Thank you very much! I will try it. After modifying it, recompiling, and then using export, will there be a debug file?

mive93 commented 3 years ago

A debug folder. Just create the debug and layers folder in advance.

B1ackg0d commented 3 years ago

Thank you for your reply! I have another question. Do the parameters in the training cfg, the exported cfg and the cfg used by tkdnn all need to be consistent?Especially the size of batchsize.

mive93 commented 3 years ago

Yes, they need to be consistent, but the batch size of that config (only used for training) is not used in tkDNN.

B1ackg0d commented 3 years ago

It's quite easy, just leave GPU=0 in darknet makefile.

If I use VS2019 to compile darknet in Windows, can I just use this sln directly? image