These are chat archives for beniz/deepdetect

18th
May 2016
Tim Patrick
@sudoawesomeness
May 18 2016 06:16
thanks a lot @beniz, I will switch to GPU soon. I'm feeling the need to tune the parameters. Is there any way to gather the train_loss data for each iteration to visualise the trend?
Emmanuel Benazera
@beniz
May 18 2016 07:57
@sudoawesomeness have you looked at the ...measure_hist API parameter ? You can then visualize the trends with whatever tool
A more powerful tool is https://github.com/Isaacpm/dboard that uses elasticsearch and kibana to record and visalize all metrics.
Tim Patrick
@sudoawesomeness
May 18 2016 10:45
@beniz I only changed the call for getting info of training job as:
curl -X GET "http://localhost:8080/train?service=l0resnet&job=1&parameters.output.measure_hist=true"
getting "body":{"measure_hist":{},"measure":{"iteration":2.0,"train_loss":0.30396172404289248}}}
output in training POST call:
\"output\":{\"measure\":[\"acc\",\"mcll\",\"f1\"]}
Emmanuel Benazera
@beniz
May 18 2016 11:07
meaures are computed every test_interval iterations against your test set. If you are unsure what this means, dd doesn't abstract you completely from knowing the ML basics.
Emmanuel Benazera
@beniz
May 18 2016 12:35
@sudoawesomeness let me know if something is not clear.
Emmanuel Benazera
@beniz
May 18 2016 18:49
@sudoawesomeness how did you accomodate your first trial on OSX ? Did you get onto a different platform ?