{ "id": 18935, "title": "Deep learning tutorial error on kernel size, kernel_h and kernel_w", "url": "/competitions/second-annual-data-science-bowl/discussion/18935", "author_name": "", "post_date": "2016-02-13T08:07:21.727Z", "votes": null, "comment_count": 7, "views": 1499, "content": "

I'm new to Caffe, and I'm following Deep Learning Tutorial.

\n\n

I have created a python script to read the images, and create the databases (Steps 1 and 2). It works. Then I created another script to train and validate (steps 3-15) and also solver file. \nThe script creates the networks (fcn_train and fcn_test, contents match tutorial), however the script fails at \nsolver = caffe.SGDSolver('fcn_solver.prototxt")\nThe error message is below:

\n\n

I0213 01:30:33.690608 24127 layer_factory.hpp:76] Creating layer data

\n\n

I0213 01:30:33.691711 24127 net.cpp:111] Creating Layer data

\n\n

I0213 01:30:33.691727 24127 net.cpp:434] data -> data

\n\n

I0213 01:30:33.69281124379 db_lmdb.cpp:22] Opened lmdb train_images_lmdb/

\n\n

.....

\n\n

I0213 01:30:33.744316 24127 net.cpp:478] upscore <- score_classes

\n\n

I0213 01:30:33.744321 24127 net.cpp:434] upscore -> upscore

\n\n

F0213 01:30:33.744328 24127 base_conv_layer.cpp:19] Check failed: !conv_param.has_kernel_size() != !(conv_param.has_kernel_h() && conv_param.has_kernel_w()) Filter size is kernel_size OR kernel_h and kernel_w; not both * Check failure stack trace: *

\n\n

In fcn_ files, there is only a reference to kernel_size, and no other reference to kernel_h or kernel_w.

\n\n

Does anybody else have the same problem, and is there a solution?

\n\n

Best,\nOB

", "messages": [ { "id": "107866", "postDate": "02/13/2016 08:07:21", "content": "

I'm new to Caffe, and I'm following Deep Learning Tutorial.

\n\n

I have created a python script to read the images, and create the databases (Steps 1 and 2). It works. Then I created another script to train and validate (steps 3-15) and also solver file. \nThe script creates the networks (fcn_train and fcn_test, contents match tutorial), however the script fails at \nsolver = caffe.SGDSolver('fcn_solver.prototxt")\nThe error message is below:

\n\n

I0213 01:30:33.690608 24127 layer_factory.hpp:76] Creating layer data

\n\n

I0213 01:30:33.691711 24127 net.cpp:111] Creating Layer data

\n\n

I0213 01:30:33.691727 24127 net.cpp:434] data -> data

\n\n

I0213 01:30:33.69281124379 db_lmdb.cpp:22] Opened lmdb train_images_lmdb/

\n\n

.....

\n\n

I0213 01:30:33.744316 24127 net.cpp:478] upscore <- score_classes

\n\n

I0213 01:30:33.744321 24127 net.cpp:434] upscore -> upscore

\n\n

F0213 01:30:33.744328 24127 base_conv_layer.cpp:19] Check failed: !conv_param.has_kernel_size() != !(conv_param.has_kernel_h() && conv_param.has_kernel_w()) Filter size is kernel_size OR kernel_h and kernel_w; not both * Check failure stack trace: *

\n\n

In fcn_ files, there is only a reference to kernel_size, and no other reference to kernel_h or kernel_w.

\n\n

Does anybody else have the same problem, and is there a solution?

\n\n

Best,\nOB

", "rawMarkdown": "I'm new to Caffe, and I'm following Deep Learning Tutorial. \r\n\r\nI have created a python script to read the images, and create the databases (Steps 1 and 2). It works. Then I created another script to train and validate (steps 3-15) and also solver file. \r\nThe script creates the networks (fcn_train and fcn_test, contents match tutorial), however the script fails at \r\nsolver = caffe.SGDSolver('fcn_solver.prototxt\")\r\nThe error message is below:\r\n\r\nI0213 01:30:33.690608 24127 layer_factory.hpp:76] Creating layer data\r\n\r\nI0213 01:30:33.691711 24127 net.cpp:111] Creating Layer data \r\n\r\nI0213 01:30:33.691727 24127 net.cpp:434] data -> data \r\n\r\nI0213 01:30:33.69281124379 db_lmdb.cpp:22] Opened lmdb train_images_lmdb/\r\n\r\n ..... \r\n\r\nI0213 01:30:33.744316 24127 net.cpp:478] upscore <- score_classes \r\n\r\nI0213 01:30:33.744321 24127 net.cpp:434] upscore -> upscore \r\n\r\nF0213 01:30:33.744328 24127 base_conv_layer.cpp:19] Check failed: !conv_param.has_kernel_size() != !(conv_param.has_kernel_h() && conv_param.has_kernel_w()) Filter size is kernel_size OR kernel_h and kernel_w; not both *** Check failure stack trace: ***\r\n\r\nIn fcn_ files, there is only a reference to kernel_size, and no other reference to kernel_h or kernel_w.\r\n\r\nDoes anybody else have the same problem, and is there a solution?\r\n\r\nBest,\r\nOB", "votes": null }, { "id": "107885", "postDate": "02/13/2016 13:50:44", "content": "

Are you using the last Caffe version ?

", "rawMarkdown": "Are you using the last Caffe version ?", "votes": null }, { "id": "107887", "postDate": "02/13/2016 14:07:40", "content": "

Are you using exactly the network prototxt files from the tutorial or is your script creating them?

", "rawMarkdown": "Are you using exactly the network prototxt files from the tutorial or is your script creating them?", "votes": null }, { "id": "107893", "postDate": "02/13/2016 15:42:14", "content": "

@Alvaro Osvaldo,\nI was using the branch pointed at the tutorial, git clone https://github.com/longjon/caffe.git. I will compile the official master to see if it helps (the related cpp file seems to be similar though).

\n\n

@senecaur,\nMy script generates the scripts, but I have compared the generated one to the tutorials' example and they match each other perfectly.

", "rawMarkdown": "Alvaro Osvaldo,\r\nI was using the branch pointed at the tutorial, git clone https://github.com/longjon/caffe.git. I will compile the official master to see if it helps (the related cpp file seems to be similar though).\r\n\r\n@senecaur,\r\nMy script generates the scripts, but I have compared the generated one to the tutorials' example and they match each other perfectly.", "votes": null }, { "id": "107897", "postDate": "02/13/2016 16:17:17", "content": "

Obben, i don't have these problem. But you should try to use the last version.\nIt can see a bug that only happens in your operation system.

", "rawMarkdown": "Obben, i don't have these problem. But you should try to use the last version.\r\nIt can see a bug that only happens in your operation system.", "votes": null }, { "id": "107911", "postDate": "02/13/2016 18:49:12", "content": "

Thanks Alvaro,\nThe latest version on the hub gives a similar but slightly different error

\n\n

F0213 12:46:52.871088 9270 base_conv_layer.cpp:36] Check failed: num_kernel_dims == 1 || num_kernel_dims == num_spatial_axes_ kernel_size must be specified once, or once per spatial dimension (kernel_size specified 0 times; 2 spatial dims).

\n\n

I will keep investigating the root cause .

\n\n

edit:I suspect it is related to deconvolution changes. I made them offline, but of course everytime I run the script it regenerates the network files. I will try and update.

", "rawMarkdown": "Thanks Alvaro,\r\nThe latest version on the hub gives a similar but slightly different error\r\n\r\nF0213 12:46:52.871088 9270 base_conv_layer.cpp:36] Check failed: num_kernel_dims == 1 || num_kernel_dims == num_spatial_axes_ kernel_size must be specified once, or once per spatial dimension (kernel_size specified 0 times; 2 spatial dims).\r\n\r\n\r\nI will keep investigating the root cause .\r\n\r\nedit:I suspect it is related to deconvolution changes. I made them offline, but of course everytime I run the script it regenerates the network files. I will try and update.", "votes": null }, { "id": "107916", "postDate": "02/13/2016 19:00:27", "content": "

Yes, that was it.\nThe latest caffe doesn't work, but the one from the tutorial works after I created the networks once and manually changed them.\nNow it is running albeit slowly. Most likely I will recompile caffe with cuNN support.

", "rawMarkdown": "Yes, that was it.\r\nThe latest caffe doesn't work, but the one from the tutorial works after I created the networks once and manually changed them.\r\nNow it is running albeit slowly. Most likely I will recompile caffe with cuNN support.", "votes": null }, { "id": "107980", "postDate": "02/14/2016 13:22:38", "content": "

That complicated O.o\nGood luck Obben. :D

", "rawMarkdown": "That complicated O.o\r\nGood luck Obben. :D", "votes": null } ], "comments": [ { "id": 107885, "author_name": "alvaroosvaldo", "author_url": "", "post_date": "02/13/2016 13:50:44", "content": "

Are you using the last Caffe version ?

", "votes": null, "replies": [] }, { "id": 107887, "author_name": "senecaur", "author_url": "", "post_date": "02/13/2016 14:07:40", "content": "

Are you using exactly the network prototxt files from the tutorial or is your script creating them?

", "votes": null, "replies": [] }, { "id": 107893, "author_name": "obbenb", "author_url": "", "post_date": "02/13/2016 15:42:14", "content": "

@Alvaro Osvaldo,\nI was using the branch pointed at the tutorial, git clone https://github.com/longjon/caffe.git. I will compile the official master to see if it helps (the related cpp file seems to be similar though).

\n\n

@senecaur,\nMy script generates the scripts, but I have compared the generated one to the tutorials' example and they match each other perfectly.

", "votes": null, "replies": [] }, { "id": 107897, "author_name": "alvaroosvaldo", "author_url": "", "post_date": "02/13/2016 16:17:17", "content": "

Obben, i don't have these problem. But you should try to use the last version.\nIt can see a bug that only happens in your operation system.

", "votes": null, "replies": [] }, { "id": 107911, "author_name": "obbenb", "author_url": "", "post_date": "02/13/2016 18:49:12", "content": "

Thanks Alvaro,\nThe latest version on the hub gives a similar but slightly different error

\n\n

F0213 12:46:52.871088 9270 base_conv_layer.cpp:36] Check failed: num_kernel_dims == 1 || num_kernel_dims == num_spatial_axes_ kernel_size must be specified once, or once per spatial dimension (kernel_size specified 0 times; 2 spatial dims).

\n\n

I will keep investigating the root cause .

\n\n

edit:I suspect it is related to deconvolution changes. I made them offline, but of course everytime I run the script it regenerates the network files. I will try and update.

", "votes": null, "replies": [] }, { "id": 107916, "author_name": "obbenb", "author_url": "", "post_date": "02/13/2016 19:00:27", "content": "

Yes, that was it.\nThe latest caffe doesn't work, but the one from the tutorial works after I created the networks once and manually changed them.\nNow it is running albeit slowly. Most likely I will recompile caffe with cuNN support.

", "votes": null, "replies": [] }, { "id": 107980, "author_name": "alvaroosvaldo", "author_url": "", "post_date": "02/14/2016 13:22:38", "content": "

That complicated O.o\nGood luck Obben. :D

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "107866": "I'm new to Caffe, and I'm following Deep Learning Tutorial. \r\n\r\nI have created a python script to read the images, and create the databases (Steps 1 and 2). It works. Then I created another script to train and validate (steps 3-15) and also solver file. \r\nThe script creates the networks (fcn_train and fcn_test, contents match tutorial), however the script fails at \r\nsolver = caffe.SGDSolver('fcn_solver.prototxt\")\r\nThe error message is below:\r\n\r\nI0213 01:30:33.690608 24127 layer_factory.hpp:76] Creating layer data\r\n\r\nI0213 01:30:33.691711 24127 net.cpp:111] Creating Layer data \r\n\r\nI0213 01:30:33.691727 24127 net.cpp:434] data -> data \r\n\r\nI0213 01:30:33.69281124379 db_lmdb.cpp:22] Opened lmdb train_images_lmdb/\r\n\r\n ..... \r\n\r\nI0213 01:30:33.744316 24127 net.cpp:478] upscore <- score_classes \r\n\r\nI0213 01:30:33.744321 24127 net.cpp:434] upscore -> upscore \r\n\r\nF0213 01:30:33.744328 24127 base_conv_layer.cpp:19] Check failed: !conv_param.has_kernel_size() != !(conv_param.has_kernel_h() && conv_param.has_kernel_w()) Filter size is kernel_size OR kernel_h and kernel_w; not both *** Check failure stack trace: ***\r\n\r\nIn fcn_ files, there is only a reference to kernel_size, and no other reference to kernel_h or kernel_w.\r\n\r\nDoes anybody else have the same problem, and is there a solution?\r\n\r\nBest,\r\nOB", "107885": "Are you using the last Caffe version ?", "107887": "Are you using exactly the network prototxt files from the tutorial or is your script creating them?", "107893": "Alvaro Osvaldo,\r\nI was using the branch pointed at the tutorial, git clone https://github.com/longjon/caffe.git. I will compile the official master to see if it helps (the related cpp file seems to be similar though).\r\n\r\n@senecaur,\r\nMy script generates the scripts, but I have compared the generated one to the tutorials' example and they match each other perfectly.", "107897": "Obben, i don't have these problem. But you should try to use the last version.\r\nIt can see a bug that only happens in your operation system.", "107911": "Thanks Alvaro,\r\nThe latest version on the hub gives a similar but slightly different error\r\n\r\nF0213 12:46:52.871088 9270 base_conv_layer.cpp:36] Check failed: num_kernel_dims == 1 || num_kernel_dims == num_spatial_axes_ kernel_size must be specified once, or once per spatial dimension (kernel_size specified 0 times; 2 spatial dims).\r\n\r\n\r\nI will keep investigating the root cause .\r\n\r\nedit:I suspect it is related to deconvolution changes. I made them offline, but of course everytime I run the script it regenerates the network files. I will try and update.", "107916": "Yes, that was it.\r\nThe latest caffe doesn't work, but the one from the tutorial works after I created the networks once and manually changed them.\r\nNow it is running albeit slowly. Most likely I will recompile caffe with cuNN support.", "107980": "That complicated O.o\r\nGood luck Obben. :D" }, "source": "meta" }