search

Vladkryvoruchko / PSPNet-Keras-tensorflow

TensorFlow implementation of original paper : https://github.com/hszhao/PSPNet
MIT License
394 stars 174 forks source link

error in pspnet.prototxt #8

Open jeansely opened 7 years ago

jeansely commented 7 years ago

Hello, when run weight-converter this error occur:

"caffe.LayerParameter" has no field named "bn_param".

                      i replace it with "batch_norm_param"

again occur below error: "caffe.BatchNormParameter" has no field named "slope_filler".

Vladkryvoruchko commented 7 years ago

Hello @jeansely . There is modified prototxt in /utils/model

jeansely commented 7 years ago

now in /utils/model the prototxt is ; bn_param { slope_filler { type: "constant" value: 1 } bias_filler { type: "constant" value: 0 } frozen: true momentum: 0.95

what are replace these lines ?

jefequien commented 7 years ago

It is from here. https://github.com/hszhao/PSPNet/blob/master/evaluation/prototxt/pspnet50_ADE20K_473.prototxt

jeansely commented 7 years ago

when run weigth -converter occur error in prototxt "line 109"

has no field named "bn_param". and "caffe.BatchNormParameter" has no field named "slope_filler".

CuriousCat-7 commented 6 years ago

I hold the same problem

tangsanli5201 commented 6 years ago

I hold the same problem too.

Vladkryvoruchko commented 6 years ago

@CuriousCat-7 @tangsanli5201 Please use already converted h5, json weights. Links that can be found in README.md I just cannot remember what was done with this converter, so if I find out i will leave a comment in this thread

tangsanli5201 commented 6 years ago

@Vladkryvoruchko Yeah. I tried to use the already converted json weights, but it raised an error: 'bad marshal data...'., which I thought it is caused by the different version of Keras( mine is 2.1.2). And it's solved now: I generated the json file through Keras 2.1.2 and replaced the old one. It's working. Thanks!

RahavLussato commented 6 years ago

@tangsanli5201 how did you manage to generate the json ? i have the same bad marshal data... issue