These are chat archives for beniz/deepdetect

21st
Mar 2017
Gilbow
@Gilbow
Mar 21 2017 10:54
Hi, i just finish my first training (ssd), and i have some trouble to add my now caffemodel to deepdetect.
model seems to work fine with standalone python script
but deepdetect give me this message : The NetState phase (0) differed from the phase (1) specified by a rule in layer detection_out
Emmanuel Benazera
@beniz
Mar 21 2017 10:56
look at https://github.com/beniz/deepdetect/tree/master/demo/objdetect and make sure your prototxt file looks similar.
Gilbow
@Gilbow
Mar 21 2017 10:56
is there some thing to change in deploy.prototxt (i quess "phase : caffe.TEST" in include directive of detection_out layer need some changes)
i already look at this
Emmanuel Benazera
@beniz
Mar 21 2017 10:58
can you run the demo ?
Gilbow
@Gilbow
Mar 21 2017 10:58
in my deploy.prototxt, in detection_out layer, i have an additionnal directive : save_output_param {
yes, all demos run fine
maybe json service configuiration need some work
about my deploy.prototxt, i try to remove save_output_param without success
i check in python caffe module caffe.TEST = 1 and caffe.TRAIN = 0
i'm sure these values stored in caffemodel/prototxt and need to be 'synchronised'
Emmanuel Benazera
@beniz
Mar 21 2017 11:02
if you've trained from standard SSD+VGG the deploy.prototxt should be very similar or the same as for the VOC demo, but for the number of classes. If the VOC demo works, then your prototxt differs.
Gilbow
@Gilbow
Mar 21 2017 11:03
let's diff
you are right
Gilbow
@Gilbow
Mar 21 2017 11:09
they are different. mainly values,
but first layer looks like this in my deploy
input: "data"
input_shape { ...
[then batch_size and input data dimension]
formating as follow : dim: 1 dim: 3 dim: 300 dim: 300
some layers are added (conv9*)
Gilbow
@Gilbow
Mar 21 2017 11:24
there is the full diff between my deploy (left one) and obj detect demo (right). http://pastebin.com/msxZTVft
i have to go, enjoy your meal, and thx for help :)
Gilbow
@Gilbow
Mar 21 2017 12:59
the two configurations for deepdetect services ares the sames (except for model location.)
Emmanuel Benazera
@beniz
Mar 21 2017 13:00
are you sure ?
Gilbow
@Gilbow
Mar 21 2017 13:00
yes
Emmanuel Benazera
@beniz
Mar 21 2017 13:00
certain ?
Gilbow
@Gilbow
Mar 21 2017 13:01
diff folder/192.168.1.231/search.json folder/192.168.1.231/plateface.json
5c5

< "repository": "/home/imajing/imajing-dd/models/model_search"

"repository": "/home/imajing/imajing-dd/models/model_plateface"
search is the demo
Emmanuel Benazera
@beniz
Mar 21 2017 13:01
that's not what your pastebin is saying...
Gilbow
@Gilbow
Mar 21 2017 13:01
plateface is my model
i have other issues
Emmanuel Benazera
@beniz
Mar 21 2017 13:03
look at your pastebin again, and read the FAQ. Sometimes computer science looks like a Zelda quest, but that's the only way to truely become indepedent right ? ^^
Gilbow
@Gilbow
Mar 21 2017 13:04
when i try ssd_detect script with a caffe from github, model load correctly and find a label woth a bbox.
when i use caffe_dd from deepdetect build, i have this error : RuntimeError: ./include/caffe/llogging.h:153 / Fatal Caffe error
this is a python error, : net = caffe.Net(model_def, model_weights, caffe.TEST)
i never play Zelda, only SC2 :)
but i see what you mean
Emmanuel Benazera
@beniz
Mar 21 2017 13:08
read the faq on how to import a model, then move forward to the next error if any.
Gilbow
@Gilbow
Mar 21 2017 13:08
i'm sorry, i really don't know what i have to find
ok, i will search in fac.
faq
Emmanuel Benazera
@beniz
Mar 21 2017 13:09
have you fixed your prototxt according to the diff in your pastebin??
Gilbow
@Gilbow
Mar 21 2017 13:10
i don't know what and how to fix :(
i take a look at the faq, and come back after :)
Gilbow
@Gilbow
Mar 21 2017 13:18
OK, i got it.
let's try !
Gilbow
@Gilbow
Mar 21 2017 14:23
COOL !! my trained model is :
1) loaded in deepdetect (after some ajustment in service configuration, and some changes in prototxt explained un the FAQ)
2) and return correct boxes :D
Thx a lot @beniz
Emmanuel Benazera
@beniz
Mar 21 2017 15:40
w00t
:)
Mona Jalal
@monajalal
Mar 21 2017 20:05
Hi, has anyone used DeepDetect for detecting human? is there any walkthrough or pretrained weight for so?
Emmanuel Benazera
@beniz
Mar 21 2017 20:29
We do yes, faces and persons via the object detector. We don't provide these models though.
Look caffe SSD up and try to find a dataset unless you have your own data