These are chat archives for beniz/deepdetect

27th
Oct 2017
rperdon
@rperdon
Oct 27 2017 14:33
My alexnet model exported from digits is 200mb in size
Emmanuel Benazera
@beniz
Oct 27 2017 14:35
that's OK, but I would recommend a GoogleNet. There's almost no reason to use AlexNet
rperdon
@rperdon
Oct 27 2017 14:36
I started uploading that model
Emmanuel Benazera
@beniz
Oct 27 2017 14:36
OK
rperdon
@rperdon
Oct 27 2017 14:36
I'll train some new models off googlenet
its uploading to dropbox now
for our purposes, i think it should highlight the differences in the input and outpout
In the end, I'm hoping to get back to training models and testing for classification accuracy while simultaneously pushing development of our classification tool
Emmanuel Benazera
@beniz
Oct 27 2017 14:39
which you should do with DD directly ;)
rperdon
@rperdon
Oct 27 2017 14:39
I think you can appreciate the idea of a filter that can alleviate some of our workloads
While I can probably train them in DD, I also need to think of the partners we are working with developing models in other environments
the hope is that perhaps we can aggregate all the models despite the machine learning libraries used into something useful
The model is uploaded to dropbox, I hope you can access it and that I put in the correct e-mail address
My early testing with googlenet models did not yield the same confidence numbers as the alexnet models we made
rperdon
@rperdon
Oct 27 2017 14:44
On the supplied image, I have our classify.py reporting a 99.2560% classification
for non anime
in both OpenCV (BGR2RGB) and PIL
Emmanuel Benazera
@beniz
Oct 27 2017 17:59
I haven't received a Dropbox link. The day is over here but paste it in PM here on gitter and we'll take a look soon.
rperdon
@rperdon
Oct 27 2017 18:19
I thought it e-mailed it to you
sorry about that
I appreciate anything you can find to help on this