jmtatsch

@wtliao Unfortunately, I did not (yet) train these weights myself. Sliced/sliding prediction has so much more details because the weights were trained on 714x714 crops from the full resolution image...

@scenarios Good catch, I will fix this in #30 and reevaluate.

No, But some train code was merged recently, so someone should have. > Am 24.02.2018 um 10:54 schrieb shipeng : > > @jmtatsch.Did you train PSPnet on any dataset? >...

Just convert the npy weights with your desired keras/python Version. You’ll be fine.

just search the issues you may find something.

https://www.dropbox.com/s/slzgd07ijcyl45b/pspnet50_ade20k.npy?dl=0 https://www.dropbox.com/s/gdtjzjkox999xpg/pspnet101_cityscapes.npy?dl=0 https://www.dropbox.com/s/ff6hdkhxz8h5si1/pspnet101_voc2012.npy?dl=0

Apparently this issue is caused by the marshal serial format having changed a bit between Python 3.5 and 3.6 as discovered in https://github.com/keras-team/keras/issues/9595 . Converting the weights from numpy with...

Thought I have the same issue. @alandtse last comment helped me. I extracted my token via TeslaPy and always took the wrong too short refresh token. Now it works.

There is this https://github.com/syreel/philips-airpurifier

Looked into the code and found that there is a keep alive implemented, just the calculation of the duration is wrong. Clients should request a new SecurityToken after 75 %...