These are chat archives for beniz/deepdetect

10th
Feb 2018
going to run again with extract_layer = "pool3/3x3_s2"
my images are 300x300 per your earlier recommendation too
Emmanuel Benazera
@beniz
Feb 10 2018 10:23
Hi, thanks, you can PR and we'll look at it.
dgtlmoon
@dgtlmoon
Feb 10 2018 10:32
done
Emmanuel Benazera
@beniz
Feb 10 2018 10:39
You don't need to resize images, it's done by dd
dgtlmoon
@dgtlmoon
Feb 10 2018 10:41
i thought so :)
waiting for the pool3/3x3_s2 index to run
dgtlmoon
@dgtlmoon
Feb 10 2018 10:53
cool, so the results are MUCH better at pool3/3x3_s2, from here I'll do some better/wider testing and setup a bit of a test-bench
would be very interesting to know if training my own model from the existing data set of 200,000 images (in about 5,000 categories) would help
Emmanuel Benazera
@beniz
Feb 10 2018 11:00
It may
dgtlmoon
@dgtlmoon
Feb 10 2018 11:09
@beniz is that something better to engage your company todo?
Emmanuel Benazera
@beniz
Feb 10 2018 15:15
@dgtlmoon our daily operations are mostly with very large corps. You can send a description of your project's goals to contact@jolibrain.com including expected results, sample pictures, time constraints, operational constraints and budget.
dgtlmoon
@dgtlmoon
Feb 10 2018 15:38
@beniz ok fantastic, hopefully that step is not far off :)
is this valid? to give HTTPS URI? classif = dd.post_predict(sname,"https://mysite.com/foobar.jpg"...) ? I see that it talks about URI but its not clear if that's valid or not
I get [15:37:36] /opt/deepdetect/src/services.h:489: service imgserv mllib bad param: missing data so probably it is not...
Could easily fetch it and add some data to the shelve anyway
dgtlmoon
@dgtlmoon
Feb 10 2018 17:08
so index.ann is 340Megabytes, for 750 images.. any strategies here? that's half a megabyte per image, and the image is 30kb :D