ReadProtoFromBinaryFile problem

[likesiwell]

hello fyu,
I am using code for some tests. I download the model following your instructions.But where I run the
python predict.py images/dog.jpg --gpu 0
I always meet this problem.

upgrade_proto.cpp:86] Check failed: ReadProtoFromBinaryFile(param_file, param) Failed to parse NetParameter file: ./pretrained/dilated_convolution_context_coco.caffemodel

So, is there something wrong with you caffemodel? Or there exist some problem else?

Thank you~

[fyu]

There is nothing wrong with my caffe model. You may need to update your caffe.

Hi, I'm having the same issue, PASCAL works, but on all other datasets, it gave error on parsing, for example:
Check failed: ReadProtoFromBinaryFile(param_file, param) Failed to parse NetParameter file: pretrained/dilation10_cityscapes.caffemodel
Do you have any ideas? Thanks!

[fyu]

@AIConstructor @wenjieuoft Which commit of caffe do you use? Can you try this fork?

I was on the newest commit, i.e. f623d04, only pascal pretrained model works.
revert to older one 08c5df and it works on all datasets.

[fyu]

@wenjieuoft Thanks for trying different caffe commit. I will check the problem.

[ashafaei]

I confirm the .caffemodel files are not compatible with the current version of Caffe. Here's what worked for me.

Load the .caffemodel with the latest compatible commit. It will upgrade the model automatically. Now save the new .caffemodel. The new model file is compatible with the current version of Caffe, but it would still complain about incompatibility (at least it loads it now, and it works). If you save it again with the current commit, you will get the clean .caffemodel.

[fyu]

@ashafaei Thanks for the investigation. Can you share the compatible caffemodel with me?

[ashafaei]

@fyu Yeah, sure. I'll send you an email with the links.

[mustafaIhssan]

@ashafaei I had the same problem, and I solved it by downloading the caffemodel with wget.
apparently the sh file doesn't download it correctly or fully

[yujianwang]

@ashafaei Could you please send me the compatible Cityscapes caffe model as well? I am using the current commit of caffe.

[ashafaei]

@fyu @yujianwang
I sent these links to @fyu a few days ago; I guess he's busy and hasn't had a chance to update the files.
Here are my models that are compatible with the current commit.

http://www.cs.ubc.ca/~shafaei/dataset/d10_cs.caffemodel
http://www.cs.ubc.ca/~shafaei/dataset/d8_pascal.caffemodel
http://www.cs.ubc.ca/~shafaei/dataset/d8_camvid.caffemodel
http://www.cs.ubc.ca/~shafaei/dataset/d7_kitti.caffemodel

[Timo-hab]

Hi, i have the same issue. Only the Pascal dataset works.
When i try the caffemodel that ashafaei postet, i get:

F0530 18:35:03.448803 4698 syncedmem.cpp:56] Check failed: error == cudaSuccess (2 vs. 0) out of memory
*** Check failure stack trace: ***
Aborted (core dumped)

I logged the GPU memory and it exceeds actually 4 GiB (I use a GTX 980).
I tried to change the mini_batch in predict.py to 1, but it has not changed the error.
Does anyone have an idea, how do i get the code working?

[ashafaei]

@Timo-hab Unfortunately the larger architectures require about 6.2 GB of GPU memory.

[fyu]

Thank @ashafaei for providing the updated models! I have updated the models on my server. They are the same with the models of @ashafaei. I have tested they can be used by the old caffe code.