These are chat archives for beniz/deepdetect

6th
Apr 2016
Emmanuel Benazera
@beniz
Apr 06 2016 07:37
Hi @Isaacpm the mllib in parameters are the ones that differ between using caffe or xgboost, so it is possible the number of iterations is not passed properly. Can you post one of your calls here ?
Isaacpm
@Isaacpm
Apr 06 2016 08:54
Hey @beniz , this is the rest one:
curl -X POST "http://localhost:8080/train" -d "{\"service\":\"covert\",\"async\":true,\"parameters\":{\"mllib\":{\"gpu\":true,\"solver\":{\"iterations\":10000,\"test_interval\":200,\"base_lr\":0.05,\"solver_type\":\"ADAM\"},\"net\":{\"batch_size\":300}},\"input\":{\"shuffle\":true,\"test_split\":0.2,\"min_count\":2,\"min_word_length\":2,\"count\":false, \"tfidf\":"true"},\"output\":{\"measure\":[\"mcll\",\"f1\",\"cmdiag\",\"cmfull\"]}},\"data\":[\"/var/models_xgb/dataset\"]}"
Emmanuel Benazera
@beniz
Apr 06 2016 08:58
yhis is a post for neural nets via Caffe
what is the one you are using with xgboost ?
see examples in beniz/deepdetect#62
as you know it is in final beta, so the doc is not yet online
but it should be easy
Isaacpm
@Isaacpm
Apr 06 2016 09:01
I had to change a few things because the example is for csv and I'm suing text
using
I guess I left too much of the caffe example
Emmanuel Benazera
@beniz
Apr 06 2016 09:02
it is actually simpler in the xgb example
don't change input and output connectors
only the mllib section
Isaacpm
@Isaacpm
Apr 06 2016 09:04
so I get the mllib from the xgboost example and replace it in the caffe one, leaving input and output as they are
right?
\"tfidf\":"true", this parameter goes into the input or into the mllib section?
Emmanuel Benazera
@beniz
Apr 06 2016 09:05
correct, typically something like \"mllib\":{\"iterations\":100,\"objective\":\"multi:softprob\"}
Isaacpm
@Isaacpm
Apr 06 2016 09:05
ok, cool
need to sort one thing out and I'll give it a try
Isaacpm
@Isaacpm
Apr 06 2016 09:29
it's working, thanks!
I take there is no solver or anything like that here, going to have a read about the different parameters than be used for xgboost
there is no training loss either, right?
Emmanuel Benazera
@beniz
Apr 06 2016 09:43
train_loss is not wired, that's correct
you should get all the other measures though
all parameters will be explained in the documentation upon release
once thing you can now change within the txt connector is tfidf:true
Isaacpm
@Isaacpm
Apr 06 2016 10:05
cool, thanks, I left the other measures and all work as you say.
I did use it, I hope:
curl -X POST "http://localhost:8080/train" -d "{\"service\":\"covert\",\"async\":true,\"parameters\":{\"mllib\":{\"iterations\":1000,\"objective\":\"multi:softprob\"}},\"input\":{\"shuffle\":true,\"test_split\":0.2,\"min_count\":2,\"min_word_length\":2,\"count\":false, \"tfidf\":"true"},\"output\":{\"measure\":[\"mcll\",\"f1\",\"cmdiag\",\"cmfull\"]},\"data\":[\"/var/models_xgb/dataset\"]}"
Emmanuel Benazera
@beniz
Apr 06 2016 10:35
should work
Emmanuel Benazera
@beniz
Apr 06 2016 11:24
let me know
Isaacpm
@Isaacpm
Apr 06 2016 11:28
yeah, going to try shortly
sorry, I'm at my real work now, so need to do this while I wait for stuff to fnish
Emmanuel Benazera
@beniz
Apr 06 2016 11:31
oh no problem, thanks for testing
Isaacpm
@Isaacpm
Apr 06 2016 11:31
I hope this precision goes all the way up to 99%, lol
been reading about this xgboost for some time