RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1512946747676/work/torch/lib/THC/generic/THCStorage.cu:58 about deepglobe-road-extraction-challenge HOT 11 OPEN

I think it's because your memory is too small.

from deepglobe-road-extraction-challenge.

To train dlink34 on 1024*1024 images, your GPU should have at least 2G memory.

Maybe you can resize the images/masks to small size like 512*512 or crop small patches from the origin images/masks.

from deepglobe-road-extraction-challenge.

Speaking of image size: Does fixed image size have an impact on network performance? Does the fixed image length and width affect the performance of the network?When I use the 10241024 .shp file to diced the .tif file, there are often some dicing that are not 10241024. If it is not fixed, how can I modify the code?

from deepglobe-road-extraction-challenge.

The network can receive any input size, but due to the pooling layers and transposed-conv layers, the forward pass should have additional padding options. For computational convenience, better pad the input size to multiple of 32.

from deepglobe-road-extraction-challenge.

ok,thank you very much

from deepglobe-road-extraction-challenge.

ok,thank you very much

from deepglobe-road-extraction-challenge.

@ZuxianShi @zlkanata thanks very much, it's right that the memory is too small .

from deepglobe-road-extraction-challenge.

ok,thank you very much

from deepglobe-road-extraction-challenge.

Hello @zlkanata, I am a novice. I want to run your code but I meet the same problem too. I want to know how to resize the big datasets to 512*512, by what?

from deepglobe-road-extraction-challenge.

Hello,does the code run on cpu?can I test on cpu?Thank you very much.

from deepglobe-road-extraction-challenge.

Hello,does the code run on cpu?can I test on cpu?Thank you very much.

yes，you can run the code on cpu, I have tried, but It is very slow

from deepglobe-road-extraction-challenge.