Bodo Kaiser
Bodo Kaiser
There may be implementation details that require you to change tensor shapes if you use CUDA. That said I did not do any work with pytorch recently, thus there is...
Hey, yes that makes sense! Would you mind to send a PR? > On 21. Nov 2018, at 11:07, toby wrote: > > I think the most contribution in segnet...
@Viarow I guess this is a mistake then. Would you mind trying the correct version?
If I understood correctly the problem can be found with the index selectors here: self.dec1 = features[0: 4] #[0,3] self.dec2 = features[5: 9] #[5,8] self.dec3 = features[10: 16] #[10,15] self.dec4...
Hi, I added loss weights and now retraining SegNet on AWS (this time for more then 31 epochs). I post the results when they are ready. Do you have any...
Hmm okay I don't think this changes much. After 11 epochs there are not really much improvements to unweighted loss (except that background is now missing..).
Wow the couch segmentation looks really good! As far as I see the "fuse" problem is solved with: - Deconvolution (Transposed Convolution) - Upsampling - Unpooling Did you make any...
Ah good to know! Thank you very much.
So you already ran ~100 epochs training on both variants? Do you have more ideas how to improve SegNet performance? Maybe apply soft learn rates for the pretrained layers?
Is it correct to pass the [indices of the prior layer](https://github.com/ruthcfong/piwise/blob/master/piwise/network.py#L342-L346) and not the indices of the "prior-prior" layer?