I still don’t save but just kill， I uploaded my data and configuration, can you help me see if it works?
test.tar.gz (1.0 MB)
It can be found normally on the configuration and the original dataset I uploaded, but not on the dataset. I don’t know the data there. Is this because my dataset has already appeared?
I found it because of the explosion, but resnet18 training batch_size_per_gpu: 1 is already the lowest, or training
The computer’s memory is also covered with dust, and then the program kills
Can you upload your training spec file?
detectnet_v2_train_resnet18_kitti.txt (5.2 KB)
I found out last night that if I replaced it with the resnet10 model, I couldn’t train it either. It would also cause video memory and memory overflow, and the program would be killed.
In your case, training on 1070ti, please try to train a smaller model.
Change
output_image_width: 960 output_image_height: 544
to
output_image_width: 480 output_image_height: 272
Also, set below in augmentation_config
enable_auto _resize: true
After enabling the enable_auto_resize parameter, are the width and height set earlier not needed? Or is it meaningless?
It is needed.
See https://docs.nvidia.com/tao/tao-toolkit/text/object_detection/detectnet_v2.html#training-the-model
For detectnet_v2, it does not support training on images of multiple resolutions. So, firstly, make sure all the training images have the same resolution.
Then if you train a model as below and set enable_auto _resize: true
output_image_width: 480
output_image_height: 272
then, it will resize your training images to 480x272.
Thanks, I went through the docs and found this trick, I’ll try it again in the evening, thanks again!
After the image is automatically resized, will it affect the marked information, resulting in inaccurate detection?
No.
Just need to make sure your original images/labels are correct.
Can you help me take a look again?
Please run with a new result folder.
thanks for you
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.
