search

TimoSaemann / caffe-segnet-cudnn5

This repository was a fork of BVLC/caffe and includes the upsample, bn, dense_image_data and softmax_with_loss (with class weighting) layers of caffe-segnet (https://github.com/alexgkendall/caffe-segnet) to run SegNet with cuDNN version 5.
Other
176 stars 127 forks source link

Not able to reproduce results after following Segnet tutorial #14

Closed ChidanandKumarKS closed 7 years ago

ChidanandKumarKS commented 7 years ago

I have gone thru tutorials given on SEGNET from the site http://mi.eng.cam.ac.uk/projects/segnet/tutorial.html.

I did exactly the same steps for SEGNET_BASIC and BAYGNET_SEGNET_BASIC since i had only 4GB GPU. Apart from that i have given only batch size of 1 instead of 4 to avoid out of memory bound. Results of SEGNET_BASIC and BAYGNET_SEGNET_BASIC are very from ground truth. Kindly requesting you to let me know what mistake i have done so that i can get decent results compared to ground truth Regards K S Chidanand Kumar

TimoSaemann commented 7 years ago

Unfortunately the batch size matters a lot. That's probably the reason for your worse results. To improve the results anyway you could try to train it with more iterations.

nathanin commented 7 years ago

With a smaller GPU, try to set the iter_size parameter in the solver.

See this discussion from Caffe Users group