These are chat archives for beniz/deepdetect

10th
Jan 2017
Emmanuel Benazera
@beniz
Jan 10 2017 08:37
Look at https://deepdetect.com/tutorials/object-detector/, for detectors number of classes are in general the number object classes + background (no object), that is 21 in the tutorial.
Abhinav Guru
@abhiguru
Jan 10 2017 16:15
merci, In case of the example with Voc712 SSD300 the corresp.txt contains the list of 21 classes that form the model
but what if I want to add a model form caffezoo that might not be added to DeepDetect yet like VGG19
looking at deploy.prototxt does not tell me what are the number of classes
Emmanuel Benazera
@beniz
Jan 10 2017 16:18
yes it does, the final layer before the softmax usually has the number of classes as its output parameter
Abhinav Guru
@abhiguru
Jan 10 2017 18:39
Will have to check the spec on prototxt by caffe
*i think the prototxt file structure seems vary a bit
Emmanuel Benazera
@beniz
Jan 10 2017 18:58
if you'd like a model zoom model into DD, point to it here, I'll tell you whether it fits, and can help to move forward
Abhinav Guru
@abhiguru
Jan 10 2017 19:07
will do
Abhinav Guru
@abhiguru
Jan 10 2017 21:14
Trying to use VGG-19 (the summary section has model VG-16 linked instead of VGG-19) https://github.com/beniz/deepdetect, http://www.robots.ox.ac.uk/~vgg/research/very_deep/ has the model and deploy.prototxt
placed the deploy.prototxt in templates/caffe/vgg_19
and the model file in models/vgg_19/
I could not figure out from deploy.prototxt what the number of classes were based on the layers above softmax
tried searching a bit over the web and found mentions to try 1000 and 224x224
so my create looks like curl -X PUT "http://192.168.10.132:8080/services/imageserv" -d '{
"mllib":"caffe",
"description":"image classification service",
"type":"supervised",
"parameters":{
"input":{
"connector":"image",
"height":224,
"width":224
},
"mllib":{
"template":"vgg_19",
"nclasses":1000
}
},
"model":{
"templates":"/home/abhinav/ws/deepdetect/templates/caffe",
"repository":"../../models/vgg_19"
}
}'
Abhinav Guru
@abhiguru
Jan 10 2017 21:27
I was able to create vgg_16 but the predictions have cat:"0...9" instead of names
Abhinav Guru
@abhiguru
Jan 10 2017 22:00
Trying to look for correspondence.txt for vgg16 cant find it here https://deepdetect.com/models/vgg_16/
Emmanuel Benazera
@beniz
Jan 10 2017 22:04
because it's in the templates in the code repository. It s really easy to get vgg19 working, just compare to vgg16. Take a careful look at the tutorial, it'd be all there, the category correspondence file etc.
Abhinav Guru
@abhiguru
Jan 10 2017 22:08
the only correspondence file i could find was one for imagenet https://github.com/beniz/deepdetect/tree/master/templates/caffe/vgg_16 ?
Abhinav Guru
@abhiguru
Jan 10 2017 23:08
and a bunch of others ./datasets/imagenet/corresp_ilsvrc12.txt
./caffezoo/bags/corresp.txt
./caffezoo/trees/corresp.txt
./caffezoo/buildings/corresp.txt
./caffezoo/footwear/corresp.txt
./caffezoo/furnitures/corresp.txt
./caffezoo/sports/corresp.txt
./caffezoo/clothing/corresp.txt
./caffezoo/fabric/corresp.txt
did a full *.txt on the repo and looked thru the server dir