Test the KITTI 2015 dataset with low GPU or CPU

[TrackingBird]

Thanks for your working. Sorry to bother you. I just want to test one image pairs of KITTI 2015 with your pre-trained model. python submission.py --datapath /media/jennifer/Papers/1_StereoData/KITTI2015_data_scene_flow/testing/ --loadmodel pretrained_model_KITTI2015.tar While the process is out of memory. I just change the image size in submission.py, I also get some errors. I choose no-cuda, there are also some errors. Can I test kitti 2015 dataset with low GPU or CPU? Thanks.

[JiaRenChang]

@TrackingBird I am afraid not.
This model need around 4500 MB GPU memory to inference one pair of KITTI images. However, we are working on reducing this model to embedded platforms.

[tlkvstepan]

Hello, Why it takes so much memory for the inference? The bottleneck I guess is cost a array at the beginning of contraction part of size h/4w/4d/4*64.. it should be less than GB with floats if we don’t store intermediate results and gradients.

On 16 May 2018, at 14:25, Jia-Ren Chang [email protected] wrote:

@TrackingBird I am afraid not. This model need around 4500 MB GPU memory to inference one pair of KITTI images. However, we are working on reducing this model to embedded platforms.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[sky-github]

Hi, have you made any progress on reducing this model to embeded platforms? @JiaRenChang

[passion3394]

@TrackingBird please refer this https://github.com/passion3394/PSMNet_CPU

[Abasin123]

@TrackingBird please refer this https://github.com/passion3394/PSMNet_CPU

Hi，What changes have you made？