These are chat archives for beniz/deepdetect

9th
Apr 2016
Emmanuel Benazera
@beniz
Apr 09 2016 12:00
@Isaacpm should get about the same perfs I believe. What does your training call look like ?
Isaacpm
@Isaacpm
Apr 09 2016 15:09
@beniz I used this one:
curl -X POST "http://localhost:8080/train" -d "{\"service\":\"covert\",\"async\":true,\"parameters\":{\"mllib\":{\"iterations\":20000,\"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 09 2016 15:37
100 iterations would do I think. Also you can "mllib":{"iterations":100,"objective":"multi:softprob","booster_params":{"max_depth":10}}
Isaacpm
@Isaacpm
Apr 09 2016 15:53
thanks, will try that and let you know if it improves
Isaacpm
@Isaacpm
Apr 09 2016 16:30
tried with a few different depths and never goes above 0.70 precision, but I think it's fine, we get about only 7% error when we check all the training data against the model and compare the categories, and we should do better when we just add more products to the low performing categories
I'm going to try a layered approach now, and see if that improves it, but I think we are happy enough with the current performance of the best combination
Emmanuel Benazera
@beniz
Apr 09 2016 16:35
OK, nice to hear.