search

hszhao / PSPNet

Pyramid Scene Parsing Network, CVPR2017.
https://hszhao.github.io/projects/pspnet
Other
1.58k stars 544 forks source link

Convert to Tensorflow model. How to deal with Interpolation layer? Help with the Keras imlementation #45

Open Vladkryvoruchko opened 7 years ago

Vladkryvoruchko commented 7 years ago

I want to try run this model in tensorflow, and is using caffe-tensorflow converter.

There was some minor errors which was fixed, but the main trouble is the interpolation layer. If i am right - this is the custom layer made specifically for the semantic segmentation networks, and is absent in the master branch of caffe, so its not implemented in any converters.

Has anyone tried to convert it? And how did you face this problem?

Updated:

Take a look at my repo - implementation in Keras, there are some mistakes which are leading to bad results. Help highly appreciated Vladkryvoruchko/PSPNet-Keras-tensorflow

Updated:

Now it is working Keras project with good results

aneesh3108 commented 7 years ago

Hi @Vladkryvoruchko , I've been trying the same but haven't really got past an error telling me that "bn_param" is not a part of Caffe.

Can you possibly share your current conversion output?? I'm working on extending the interpolation layer from DeepLab to tensorflow, but the current code seems to be a bottleneck for me.

Thanks!

Vladkryvoruchko commented 7 years ago

Hi @aneesh3108 . I'm fine with bn_param and caffe works well. But earlier i have made some corrections to prototxt selection_021

I'm working now with Keras and things seems to be fine, but havent finished yet

Vladkryvoruchko commented 7 years ago

@aneesh3108 Please take a look at my results. But results are not satisfying and i cant figure it out why Vladkryvoruchko/PSPNet-Keras-tensorflow EDIT: Now it is working Keras project with good results

aneesh3108 commented 7 years ago

@Vladkryvoruchko On it this coming week. This is a great starter.

wtliao commented 7 years ago

@amanjhunjhunwala Hi, thanks for sharing your code. But the link seems inaccessible. Furthermore, do you have any idea how to add an Auxillary loss layer?

amanjhunjhunwala commented 7 years ago

@wtliao Due to certain reasons, I needed to close-source the code ! Sorry for the inconvenience !

jaypatravali commented 6 years ago

@Vladkryvoruchko Hey why is frozen=true in your prototxt, is this deploy. Just checking!