Darknet: run yolov3 with GPU:CUDA Error: out of memory

[zsjerongdu]

Hi,
I maked the darknet with "GPU=1,CUDNN=1,OPENCV=1" successfully，however，when I use the command "sudo ./darknet detector test cfg/coco.data cfg/yolov3.cfg yolov3.weights data/dog.jpg
",it shows:
CUDA Error: out of memory
darknet: ./src/cuda.c:36: check_error: Assertion `0' failed.
But if I use the command"sudo ./darknet detector test cfg/coco.data cfg/yolov2.cfg yolov2.weights data/dog.jpg",it can detect targets successfully.
It seems that the problem is the yolov3.What can I do to solve the problem?

[dayn9t]

I meet the same problem.
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.48 Driver Version: 390.48 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GeForce GTX 1050 Off | 00000000:09:00.0 On | N/A |
| 34% 36C P8 N/A / 75W | 309MiB / 1997MiB | 0% Default |
+-------------------------------+----------------------+----------------------+

[hjchai]

@dayn9t I think your gpu is low in memory. when Yolov3 fully loaded to gpu, it takes about 1600MB memory by default setting(416*416) on my computer, plus 300ish MiB from display and other applications, it is very like it will throw out OOM error. Try to run on a gpu with larger memory or reduce the width and height setting in your cfg file(Note: reducing the size might impact your detection results.).

[zsjerongdu]

Try to reboot and it may help.

[dayn9t]

thanks @hjchai .
I found the location of the error: darknet/src/maxpool_layer.c:46, then I found the variable 'batch' is 128.
I change the batch to 1 in cfg file, so it worked.

[aryus96]

I had the same problem with a GT740M with 4096Mo GDDR4 memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :

batch=1
subdivisions=1
width=416
height=416

[cesarhcq]

I had the same problem with a GT740M with 4096 Memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :
batch=1
subdivisions=1
width=416
height=416

That's worked for me!

Thank you very much!

[idpdka]

I had the same problem with a GT740M with 4096 Memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :
batch=1
subdivisions=1
width=416
height=416

Your solution works on me, thank you very much!!!!

[Aurora11111]

does the weight and height inference the result?

[cesarhcq]

does the weight and height inference the result?

That's a good question, i guess not!

[SonaHarutyunyan]

Here https://github.com/AlexeyAB/darknet you can find this note:
Note: if error Out of memory occurs then in .cfg-file you should increase subdivisions=16, 32 or 64

[amandazw]

I had the same problem with a GT740M with 4096 Memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :
batch=1
subdivisions=1
width=416
height=416

It works！thank you so much ！

[rvjenya]

I had the same problem with a GT740M with 4096Mo GDDR4 memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :

batch=1
subdivisions=1
width=416
height=416

Thanks. works))

[dav-ell]

Another solution that worked for me was to use one of the alternate config files: yolov3-tiny.cfg.

You'll notice that @aryus96 options is what is used in that file as well:

[net]
...
batch=1
subdivisions=1
...
width=416
height=416

So instead of using the command

./darknet detect cfg/yolov3.cfg yolov3.weights data/dog.jpg

I use

./darknet detect cfg/yolov3-tiny.cfg yolov3.weights data/dog.jpg

I can also use yolov3-openimages.cfg, yolov3-spp.cfg, and yolov3-voc.cfg without errors.

[Flock1]

I had the same problem with a GT740M with 4096Mo GDDR4 memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :

batch=1
subdivisions=1
width=416
height=416

Hi,
I'm using batch as 64 and subdivisions as 2. It still runs out of memory.

[zenogantner]

To not get an OOM during prediction, I had to set

batch=16
subdivisions=16
width=608
height=608

on my Quadro M1200 with 4GB of RAM.

[abdou31]

I had the same problem with a GT740M with 4096Mo GDDR4 memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :

batch=1
subdivisions=1
width=416
height=416

Not working , it show :
Error: You set incorrect value batch=1 for Training! You should set batch=64 subdivision=64

[zenogantner]

@abdou31 I think most of the comments inthis issue are about inference, not training.

[oshadaamila]

I had the same problem with a GT740M with 4096Mo GDDR4 memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :

batch=1
subdivisions=1
width=416
height=416

this worked in my nvidia GT940M

[abdou31]

@zenogantner but what about training?

[dhilip1]

I had the same problem with a GT740M with 4096Mo GDDR4 memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :

batch=1
subdivisions=1
width=416
height=416

It worked!! finally i am able to train my dataset!!! i thought of nvidia drivers installed already..in my laptop by default..!!

[Ingmar05]

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :

batch=1
subdivisions=1
width=416
height=416

When uncommenting batch=1 and subdivisions=1 i got it working with a detection time of 2.9 seconds.

Here https://github.com/AlexeyAB/darknet you can find this note:
Note: if error Out of memory occurs then in .cfg-file you should increase subdivisions=16, 32 or 64

But by undoing the changes and just increasing the subdivisions from 16 to 32 i got a detection time of 0.2 seconds

Running RTX 2060 with nvidia-418 and cuda 10.1

Edit:
width=608
height=608
was used throughout

[ben423423n32j14e]

I'm using a Nvidia GT 1030 (2gb memory), getting a prediction time of 0.162808 seconds with these settings:

batch=32
subdivisions=32
width=416
height=416

I have not found a settings combination to run width and height of 608, I see this error:

74 res 71 19 x 19 x1024 -> 19 x 19 x1024
CUDA Error: out of memory

Looking at nvidia-smi, it seems like it "only just" runs out of memory trying with 608, if there was an extra 500mb memory on the card I suspect it would work. :(

[twodoge]

Try to reboot and it may help.

I did what he said, and then it succeeded.

[pirate-lofy]

thanks @hjchai .
I found the location of the error: darknet/src/maxpool_layer.c:46, then I found the variable 'batch' is 128.
I change the batch to 1 in cfg file, so it worked.

I found the variable batch in the file cfg/yolov3.cfg with value 40, when I changed to 1 it produced an error saing "0 cuda nalloc failed".
but when I changed to 31 or less it worked.
I think this number will vary from a pc to another.

[AbhimanyuAryan]

I was using 2080Ti & It showed me this error. So I went through some stackoverflow links. Remove Nvidia's outdated driver(if you followed some blog post to install driver and cuda)

Download latest drivers manually from official website. You are good to go....error fixed 🎉🎉🎉

[antonmilev]

hello. I have latest Nvidia drivers and GTX 1050, nvidia-smi shows that I have 4GB GPU memory:
2019/08/14 17:19:27.728, 2400 MiB, 4096 MiB
But why darknet shows CUDA out of memory for:
batch=32
subdivisions=16
after it reaches 3488 MB? The most it is able to allocate is 2400MB, so I am forced to use:
batch=32
subdivisions=32
I wonder is it NVIDIA bug and having actually less than 4 GB usable GPU memory?

[intelltech]

Tengo una GT 740M de 2GB y funciono con yoloV3-tiny, tanto en la configuracion como los pesos.

[barzan-hayati]

Thanks @aryus96 .

I have this problem in another way. I explained my problem in Darknet docker image doesnt work after shipping to another system #4082."

First Machine: RTX 2080 Ti with 11G memory
Second Machine: Geforce 1050 Ti with 4G memory

I have downloaded a docker image with this specs:cuda9.2.148 cudnn7.5.0 opencv3.4.6 .

Then I pulled darknet and make it.I could successfully train and test tiny-yolov2 and tiny-yolov3 in docker. then I commit this docker container, saved it to a .tar file and moved it to another system. Now I want to test darknet in new system via loading docker image.

When I want to test darknet in new machine, by testing tiny-yolov2 it could not detect any objects and by testing tiny-yolov3 it failed by CUDA Error: out of memory. I think second problem rises for lower GPU memory in second machine. How should I resolve first problem?

As you mentioned, I guess second problem rises from lower GPU memory. Are you have any idea about my first problem(detection in one but could not detect in another)?

[Monkey1GIt]

I Training YOLO on VOC according to the introduction from https://pjreddie.com/darknet/yolo/#train-voc
./darknet detector train cfg/voc.data cfg/yolov3-voc.cfg darknet53.conv.74
even through I modify the cfg like:

batch=1
subdivisions=1
width=416
height=416

I also have a mistake Error: You set incorrect value batch=1 for Training! You should set batch=64 subdivision=64 like @abdou31 meet.
But when I modify the cfg to:

batch=1
subdivisions=16
width=320
height=320

Both error has disappered.
FYI: like https://github.com/pjreddie/darknet/issues/791#issuecomment-390096490 said it may impact detection precision
https://github.com/AlexeyAB/darknet#how-to-improve-object-detection also has mentioned it:

increase network resolution in your .cfg-file (height=608, width=608 or any value multiple of 32) - it will increase precision

[XunshanMan]

Hi, I used a GPU Nvidia 840M with 2GB memory, and met the same problem.
It's funny after I closed VS code, which took up around 300Mb memory, and started again, it worked. It seemed yolo need 1.3Gb memory for me.

I also tried increasing the batch and subdivision( they need to be the same, or there would be problem) in the .cfg file, it didn't work, even turned them up to 1024.

[DimasVeliz]

I had the same problem with a GT740M with 4096Mo GDDR4 memory. Nvidia 384.130, Cuda 9, CUDNN, OpenCV 3.3.

My solution to run Yolov3 perfectly was to : modify the cfg/yolov3.cfg :

batch=1
subdivisions=1
width=416
height=416

Beers and cheers for this guy! It works!

[nisarggandhewar]

How to reduce darknet training time. any cloud based service can we use, so that we can get more powerful GPU and training time can reduce.
Can we configure darknet on Google Colab.

[prh-t]

I also have a mistake Error: You set incorrect value batch=1 for Training! You should set batch=64 subdivision=64 like @abdou31 meet.
But when I modify the cfg to:

batch=1
subdivisions=16
width=320
height=320

Both error has disappered.
FYI: like https://github.com/pjreddie/darknet/issues/791#issuecomment-390096490 said it may impact detection precision
https://github.com/AlexeyAB/darknet#how-to-improve-object-detection also has mentioned it:

increase network resolution in your .cfg-file (height=608, width=608 or any value multiple of 32) - it will increase precision

Just a hypothesis: having subdivisions / batch == 16 resolves the out of memory issue... somehow...
I set mine to be batch=4, subdivisions=64, height,width = 608, 608, and it's running ok so far, fingers crossed...

[prh-t]

Try to reboot and it may help.

This might work for some as (possibly) your previous failed runs are still occupying some memory, a similar approach would be to kill all python processes.

[dhgokul]

Tried reboot device and modify yolo config file. still i am getting issue. Any help appreciated !

batch=1
subdivisions=16
width=320
height=320

batch=1
subdivisions=1
width=416
height=416

Error log :
GPU mode with 1.0 usage
2020-09-01 00:55:35.752924: W tensorflow/core/platform/profile_utils/cpu_utils.cc:98] Failed to find bogomips in /proc/cpuinfo; cannot determine CPU frequency
2020-09-01 00:55:35.755909: I tensorflow/compiler/xla/service/service.cc:161] XLA service 0x16b81140 executing computations on platform Host. Devices:
2020-09-01 00:55:35.756060: I tensorflow/compiler/xla/service/service.cc:168] StreamExecutor device (0): ,
2020-09-01 00:55:35.934491: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:965] ARM64 does not support NUMA - returning NUMA node zero
2020-09-01 00:55:35.934823: I tensorflow/compiler/xla/service/service.cc:161] XLA service 0x16b500b0 executing computations on platform CUDA. Devices:
2020-09-01 00:55:35.934900: I tensorflow/compiler/xla/service/service.cc:168] StreamExecutor device (0): NVIDIA Tegra X1, Compute Capability 5.3
2020-09-01 00:55:35.936965: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1433] Found device 0 with properties:
name: NVIDIA Tegra X1 major: 5 minor: 3 memoryClockRate(GHz): 0.9216
pciBusID: 0000:00:00.0
totalMemory: 3.87GiB freeMemory: 2.20GiB
2020-09-01 00:55:35.937042: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1512] Adding visible gpu devices: 0
2020-09-01 00:55:41.343127: I tensorflow/core/common_runtime/gpu/gpu_device.cc:984] Device interconnect StreamExecutor with strength 1 edge matrix:
2020-09-01 00:55:41.343197: I tensorflow/core/common_runtime/gpu/gpu_device.cc:990] 0
2020-09-01 00:55:41.343228: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1003] 0: N
2020-09-01 00:55:41.343471: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 3964 MB memory) -> physical GPU (device: 0, name: NVIDIA Tegra X1, pci bus id: 0000:00:00.0, compute capability: 5.3)
2020-09-01 00:55:58.584739: E tensorflow/stream_executor/cuda/cuda_driver.cc:806] failed to allocate 3.87G (4156932096 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY: out of memory
Killed