search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.77k stars 7.96k forks source link

416 x 416 Realloc error #4989

Open sunn-e opened 4 years ago

sunn-e commented 4 years ago

I am trying to build a custom model based on tiny yolo. The dataset is OIDv4 with labels converted to yolov3 format. I built darknet using visual studio 2015 on windows 10 machine . The training is in CPU mode with CUDA enabled but I'm not sure if I have enabled it. I have an nvidia 1060ti with cuda 10 and cudnn >7. The error I'm getting is realalloc. Please checkout the image. realalloc-issue-darknet

Is it possible that limited inernal storage(<6GB free) has to do anything with it? Thanks in advance.

AlexeyAB commented 4 years ago

Compile Darknet with GPU (CUDA and cuDNN) by using default darknet.sln or by using Cmake-GUI annd check CUDA and CUDNN in Cmake-GUI.

sunn-e commented 4 years ago

Thanks @AlexeyAB , I tried Cmake-GUI annd check CUDA and CUDNN in Cmake-GUI. The training has begun but it is using 100% of my cpu. I want to run it on my 1060Ti. Update: I got the problem. I forgot to make changes in makefile. Aghh. I'll come back with update and close the issue.

AlexeyAB commented 4 years ago

@sunn-e

sunn-e commented 4 years ago

@AlexeyAB ,I used darknet.exe detector train data/obj.data yolov3-tiny-obj.cfg yolov3-tiny.conv.15 -dont_show to start training. Untitledscreenshot for alexy-darknet

I did not get the output as your console. Perhaps I missed some step in compilation. I have compiled it from scratch many times. ?Used CMake GUI as per your screeenshot. I definitely have Cuda 10 and cudnn>7(my TF2-gpu and pytorch use those for training) and opencv too. For time being I switched to collab for using darknet.
Will give it another shot once I'm done with my task.

AlexeyAB commented 4 years ago

@sunn-e Do you use windows or linux? Did you do?

cmake .
make clean
make

You compiled Darknet without CUDA. You must compile Darknet with CUDA (and preferably with cuDNN) for training.