These are chat archives for beniz/deepdetect

29th
Jun 2016
Emmanuel Benazera
@beniz
Jun 29 2016 08:29
@seeb0h there are two questions in one I believe: to make your own custom convnet, either use the convnet template and define your layers via the API, or just write your own prototxt down. If you do so, it's easier to start with a basic convnet via the API and modify the prototxt files afterwards. Now, if you really just want the LeNet architecture for MNIST, take the prototxt files and modify them so that the inputs look like those of the templates (e.g. googlenet). Now, a third even easier option for MNIST, is to compile with unit tests support, and grab the MNIST files that are part of our unit tests as they include the modified prototxt. Look at the README for building with the unit tests, then look into build/examples/caffe/mnist. You can let me know how it goes.
danielgollas
@danielgollas
Jun 29 2016 20:56
@beniz Hi again! While training and cleaning up my datasets I find it useful to know the results of the testing phase, that is what probs were assigned to each image in the testdb and do they match the sample's label. Is there anyway to output that using deepdetect?
Emmanuel Benazera
@beniz
Jun 29 2016 21:02
@danielgollas good point, at the moment, the only way to do this would be run over your test.db and use the /predict call. At the moment, one hack to do so without iterating the db yourself could be to edit the deploy.prototxt file, but I wouldn't really do that. One more point, using cmdiag and/or cmfull in the measures object gives you information from the confusion matrix, which is not exactly what you are asking, but it might help.
Other than that, if you have in mind a simple way the API could be enhanced to do what you are looking for, let me know and/or open an issue.
danielgollas
@danielgollas
Jun 29 2016 21:05
Thanks for the info Emmanuel. I think using the predict approach will work. I already use the cmfull and that is exactly why I wanted to see what exactly was causing the confusion in the confusion matrix :D
I'll think about a way the API could do it. Maybe an output plugin.
danielgollas
@danielgollas
Jun 29 2016 21:21
On what I hope is a much simpler question, how can I use the python bindings to the caffe installation built by deepdetect? Is there a PYTHONPATH I can set so that my python interpreter can use it? And do I also maybe need to set some LD_LIBRARY_PATH to find the compiled libs?
Emmanuel Benazera
@beniz
Jun 29 2016 21:25
The custom caffe is available on github, you can certainly process as for the original caffe... Otherwise it is in build/caffe_dd/src/caffe_dd
danielgollas
@danielgollas
Jun 29 2016 22:58
great