These are chat archives for BVLC/caffe

19th
Aug 2015
Jaonary Rabarisoa
@jrabary
Aug 19 2015 08:33
@/all Hi everybody. Is there someone who tried to generate paired dataset for siamese network training on the fly ? In the example the dataset is generated offline. This is not scalable when the number of possible pairs is large since we may end with very huge leveldb file and I'm limited by the disk space. I'd like to generate the pairs on the fly using only the database of images. Any hint to help me to figure out how the new DataLayer, DataReader and BlockingQueue work ?
Eren Golge
@erogol
Aug 19 2015 09:09
Guys I am really fcking crazy by this error. It some times happens sometimes not in compilation.
libcaffe-nv.so: error: undefined reference to 'cv::String::allocate(unsigned long)'
I use OpenCV3 and I lnked all the libraries by LD_LIBRARY_PATH and LIBRARY_PATH
but still I have that error
any good idea?
Prasanna S
@IMG-PRCSNG
Aug 19 2015 09:12
Are you using make with more jobs ( eg make -j 8 or something )? I think the order of compilation might matter. Just guessing here.
@erogol
Eren Golge
@erogol
Aug 19 2015 09:12
yeah I use this
ok I'll try single thread
Prasanna S
@IMG-PRCSNG
Aug 19 2015 09:15
And also check if you have included "cvstd.hpp" from opencv2/core/
Eren Golge
@erogol
Aug 19 2015 09:17
you mean this right "#include "opencv2/core/cvstd.hpp" " ?
Martin Garbade
@mgarbade
Aug 19 2015 11:58
Does anyone know where the original website of the apollo module went? It had a nice example of how to use a recurrent layer and an lstm layer. Now the github account is only referencing the official caffe website
Eren Golge
@erogol
Aug 19 2015 13:48
Dear under the hood guys, do you know what makes datum.encoded() true so that is encoded image. For instance in CIFAR example, images are not encoded
Eren Golge
@erogol
Aug 19 2015 13:57
Basically I am looking for a way to scale the image size with a random scale factor in the transform pipeline. However as far as I see this is not very easy since images are considered in two ways, encoded and not encoded. Encoded images are converted to opencv Mat internally (which is a lot easier to resize) but non-encoded stays as a basic datum. I
Then I am looking a way to convert any data encoded or not into opencv format. Do you think this is the way to go or any other suggestion to do it without bells and whistles.
Prasanna S
@IMG-PRCSNG
Aug 19 2015 15:19
@erogol , I had the same problem. While creating lmdb, I set encoded = true. Check the convert_imageset.cpp in tools ( which is used in create CIFAR10 batch script ). This will slow down the pipeline as the image needs to be decoded everytime its read from the lmdb ( with ppm format, i dont think there will be a problem )but we need not implement data_transformer::Transform three times.
Eren Golge
@erogol
Aug 19 2015 15:52
I guess the most practical way is to be sure all the images written in db as encoded as you suggest