search

ShaoqingRen / SPP_net

SPP_net : Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition
363 stars 236 forks source link

Trying with eendebakpt Linux Version of Caffe #5

Closed ByzanTine closed 9 years ago

ByzanTine commented 9 years ago

I have tried with his compiled version of caffe and also the newest version of caffe. But I am getting the same error message when I run the newest version of caffe. When I run Script_spp_voc, here is aborted message.

WARNING: Logging before InitGoogleLogging() is written to STDERR [libprotobuf ERROR google/protobuf/text_format.cc:245] Error parsing text-format caffe.NetParameter: 14:14: Message type "caffe.V0LayerParameter" has no field named "conv_param". F1007 22:58:40.283839 19559 upgrade_proto.cpp:627] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: /mnt/neocortex/scratch/yejiayu/_exp/SPP_net/data/cnn_model/Zeiler_conv5/Zeiler_spm_scale224_test_conv5.prototxt

And it seems you are using V0 version proto but with some V1 features. I have tried to modify the prototxt to a V1 version, then reading the parameter will be fine but reading the weight file appears to be an error.

What should I do with this?

ShaoqingRen commented 9 years ago

the zeiler cnn net for new caffe code is shared at https://onedrive.live.com/download?resid=4006CBB8476FF777!9723&authkey=!APTWXLD_P7UN6P0&ithint=file%2czip please try this.

And the old version is not V0, so the caffe code can't read it.

ByzanTine commented 9 years ago

Can you also provide the updated version of finetune models, namely the files in the model-defs. Although it's easy to modify manually, but it's weird.

ShaoqingRen commented 9 years ago

Updated in ./model-defs

eendebakpt commented 9 years ago

The new pascal_finetune_fc_spm_solver.prototxt starts with:

The train/test net protocol buffer definition

net: "pascal_finetune_fc_spm_train_test_new.prototxt"

I changed this to

The train/test net protocol buffer definition

train_net: "pascal_finetune_fc_spm_train_test_new.prototxt" test_net: "pascal_finetune_fc_spm_train_test_new.prototxt"

to make the code work (the strings train_net and test_net are needed in Script_spp_voc.m line 102 and 103)

ShaoqingRen commented 9 years ago

Thank you very much for the reminder. I've modified the Script_spp_voc.m to support new caffe prototxt.

wistone commented 9 years ago

Hi, I download the SPP package and the corresponding caffe package a few days ago. Now I have some problem in running Script_spp_voc.m

If I use the original pascal_finetune_fc_spm_solver.prototxt, the matlab crashed with an error message of

WARNING: Logging before InitGoogleLogging() is written to STDERR I1123 15:46:26.743105 9641 common.cpp:222] Compute Capability 3.5, set cuda_num_threads = 1024 libprotobuf ERROR google/protobuf/text_format.cc:172] Error parsing text-format caffe.NetParameter: 17:23: Message type "caffe.V0LayerParameter" has no field named "inner_product_param". F1123 15:46:30.807468 9641 upgrade_proto.cpp:627] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: pascal_finetune_fc_spm_train.prototxt

If I changed to the other prototxt pascal_finetune_fc_spm_solver_new.prototxt, the matlab also crashed. Either or not I changed the prototxt following the 5th post.

So which one is the right version for training the VOC using current caffe version?

Thanks.