These are chat archives for beniz/deepdetect

9th
Oct 2016
Siddha Ganju
@sidgan
Oct 09 2016 16:47
Hi @beniz

I'm trying to fine tune AlexNet using MS COCO dataset with multilabels. These labels indicate the object category that is present in the image.

I've generated the training and validation text as follows:

path/to/image/filename.jpg 0 1 0 0 0 0 0 1 0 1 1 1

Emmanuel Benazera
@beniz
Oct 09 2016 16:47
@sidgan hello
Siddha Ganju
@sidgan
Oct 09 2016 16:48

I understand that this must be converted to an lmdb format, which I will do like this: https://github.com/BVLC/caffe/issues/1698#issuecomment-70211045

However, I am not sure if the training and validation file are in the right format. Could you please confirm if the format is right. I'm unable to find an example format online.

I found your conversation regarding the inactive and active labels in the google groups
so I thought this would be a good place to ask
Emmanuel Benazera
@beniz
Oct 09 2016 16:49
ah OK, FYI this is the channel for deepdetect, we don't really handle pure Caffe questions here
Siddha Ganju
@sidgan
Oct 09 2016 16:50
correct me if I'm wrong, deepdetect can perform multi label classification
Emmanuel Benazera
@beniz
Oct 09 2016 16:52
multi task learning is not fully automated. That being said, what are you trying to do, multi-label through regression (e.g. sigmoid cross entropy) or multi-task ?
multi-target regression is automated in DD
Siddha Ganju
@sidgan
Oct 09 2016 16:57
Multi-label through regression
Emmanuel Benazera
@beniz
Oct 09 2016 17:04
OK, image + regression. So your question is how to store since you'll slice the labels out of the lmdb, right ?
Siddha Ganju
@sidgan
Oct 09 2016 17:06
Yes
Emmanuel Benazera
@beniz
Oct 09 2016 17:06
@sidgan my hunch is that you need to store your images in straight RGB values, and add the regression labels to it
in other words, do not encode the image within the lmdb Datums
Siddha Ganju
@sidgan
Oct 09 2016 17:07
okay, let me try this out
thank you very much
Emmanuel Benazera
@beniz
Oct 09 2016 17:07
since you are using pure Caffe, you can also use two dbs
Siddha Ganju
@sidgan
Oct 09 2016 17:08
yes, I was using two
Emmanuel Benazera
@beniz
Oct 09 2016 17:09
if you dump your images values into a CSV, add the regression labels, you should be able to use DD with 'mlp' and / or 'convnet'. But as a researcher, you may not need DD, unless you don't want to convert the CSV into lmdb etc... my 2cents.