These are chat archives for beniz/deepdetect

8th
Feb 2018
dgtlmoon
@dgtlmoon
Feb 08 2018 10:51
Greetings - some general advice - If I have lots of images of tshirt's (200,000+) with different designs printed on them, I can easily classify that it's a tshirt, but what direction do I go in to find all tshirts containg a similar design to a single tshirt I've chosen? some kind of kNN problem using spotify/annoy somehow? or? thanks :)
I can probably reduce that set to 10,000 or so
Emmanuel Benazera
@beniz
Feb 08 2018 11:08
Hi, see the imgsearch (and objsearch) demos in the demo repository
dgtlmoon
@dgtlmoon
Feb 08 2018 11:09
@beniz thanks! i'll take a look
FYi I've been having a lot of luck with paperspace.com for doing my GPU work
dgtlmoon
@dgtlmoon
Feb 08 2018 15:47
Cool... getting progess
INFO - 15:44:10 - This network produces output label
INFO - 15:44:10 - Network initialization done.[15:44:10] /opt/deepdetect/src/caffelib.cc:414: Using pre-trained weights from /opt/deepdetect/demo/objsearch/model/VGG_VOC0712_SSD_300x300_iter_60000.caffemodel
INFO - 15:44:10 - Ignoring source layer mbox_loss[15:44:10] /opt/deepdetect/src/caffelib.cc:2157: Net total flops=31207751680 / total params=26064064
[15:44:10] /opt/deepdetect/src/services.h:495: service imageserv mllib bad param: unknown rois layer rois
[15:44:10] /opt/deepdetect/src/httpjsonapi.cc:394: Thu Feb  8 15:44:10 2018 UTC - ::1 "POST /predict" 400 231
Also you may want to start with imgsearchas object search is less easy to tweak
m-dennis
@m-dennis
Feb 08 2018 15:58
Hi everyone. I am new to coding. Any recommendations on getting started? AWS vs Docker
Emmanuel Benazera
@beniz
Feb 08 2018 16:06
Hi @m-dennis docker is more up-to-date than aws. Building from source gets you the last version.
m-dennis
@m-dennis
Feb 08 2018 16:07
great. any recommendations on setups to get through the tutorials?
dgtlmoon
@dgtlmoon
Feb 08 2018 16:08
@m-dennis just starting reading all the README's that were written for people to read I guess
Emmanuel Benazera
@beniz
Feb 08 2018 16:09
Training requires gpu unless you are testing a few iterations of building a very small model, inference is fine r on cpu. You may want to start with inference.
Beware that docker requires you know how to use that and share external directories
dgtlmoon
@dgtlmoon
Feb 08 2018 16:28
@beniz sorry, one more issue, getting closer.. now, no deploy file in /opt/deepdetect/demo/imgsearch/model for initializing the net
That directory contains only bvlc_googlenet.caffemodel
which is, as per the README.md what should be there
Emmanuel Benazera
@beniz
Feb 08 2018 17:04
Not sure where you are at, but using "template":"googlenet" would generate the files for you
dgtlmoon
@dgtlmoon
Feb 08 2018 17:56
aaah
Example of creating a service then listing it:

./dede --jsonapi 1 --service_name test --service_create '{"mllib":"caffe","description":"classification service","type":"supervised","parameters":{"input":{"connector":"image"},"mllib":{"template":"googlenet","nclasses":10}},"model":{"templates":"/path/to/deepdetect/templates/caffe/","repository":"/path/to/model/"}}'
Part of mllib config
@beniz Ok so.. I got the same error, after i changed the following..
--- a/demo/imgsearch/imgsearch_dd.py
-parameters_mllib = {'nclasses':nclasses}
+parameters_mllib = {'nclasses':nclasses, "template": "googlenet"}
 parameters_output = {}
 try:
I guess I need to change that runtime mllib option of dede too right?
dgtlmoon
@dgtlmoon
Feb 08 2018 18:03
I guess the imgsearch_dd.py would expect you to run dede with very custom options
dgtlmoon
@dgtlmoon
Feb 08 2018 18:10
so hmmm, I'm missing a step I feel
Emmanuel Benazera
@beniz
Feb 08 2018 18:12
It s seamless if you follow the readme I think, maybe different when using docker. Use the imgsearch.py if unsure, the built-in version requires building with USE_SIMSEARCH=ON
dgtlmoon
@dgtlmoon
Feb 08 2018 18:17
Yup, I compiled that in, even built my own Docker container with that option enabled
# https://github.com/beniz/deepdetect/tree/master/demo/imgsearch
WORKDIR /opt/deepdetect/build
RUN cmake .. -DUSE_XGBOOST=ON -DUSE_SIMSEARCH=ON
RUN make
:)
@beniz so just the default dede running with no extra mllib options? that's compatible with this imgsearch_dd.py ?
CMD ./dede -host 0.0.0.0
Emmanuel Benazera
@beniz
Feb 08 2018 18:20
Yes, the script has all the API calls that you can later replicate in any other language or framework
Use --help on the script
If you are using docker you may need to change a few directories and share one withe container.