prune weights for support yolo-v3

[zhukkang]

do you have the code for compress the yolo-v3 model, as yolo-v3 model with 416&416 need GPU memory 2~3GB, so do you have any suggest for reduce the memory?

[ArtyZe]

do you have the code for compress the yolo-v3 model, as yolo-v3 model with 416&416 need GPU memory 2~3GB, so do you have any suggest for reduce the memory?

you want to reduce the memory used in training or just test process? If you want to reduce it in forward you can commit the calloc function of updates parameters such as l.weight_updates_gpu, l.delta_gpu

another suggest is combine the convolution and batch normalization into one function, you can also save some memory

[zhukkang]

can you give some detail explain about the commit calloc function, as i just want reduce in test stage on device, not training

[ArtyZe]

can you give some detail explain about the commit calloc function, as i just want reduce in test stage on device, not training

e-mail: gaoyang917528@163.com, send me a

can you give some detail explain about the commit calloc function, as i just want reduce in test stage on device, not training

in the start of program, it will call make_convolutional_layer function, and in this function will calloc gpu memory according to your parameters, but this calloc are not all necessary for test process, so you can commit them :
l.delta_gpu = cuda_make_array(l.delta, l.batchout_hout_w*n);
if you have further question, you can send me an e-mail: gaoyang917528@163.com

[sicarioakki]

@wangnet Where did you find the pruned weights file for YOLO V2?

[AbanoubMamdouh]

Hello, I am working on a custom dataset using the tiny yolov3 architecture, I already used the AlexeyAB git repo, I got results but I need to embed prunning as well, shall I use the standard cfg file and weights file on your framework? Or there are extra steps I shall take before starting the data training?