These are chat archives for beniz/deepdetect

29th
Nov 2017
rperdon
@rperdon
Nov 29 2017 14:28
On my test material, I'm having a hard time getting anything useful out of resnet_50 within DeepDetect even after fine tuning. In testing with tensorflow and in Digits, we have been able to create some meaningful binary classifiers in a much shorter amount of time than what DD is producing. In attempting to recreate the training conditions (epochs, learning rate, momentum, resize method etc), DD is not producing anything with the same data.
Emmanuel Benazera
@beniz
Nov 29 2017 15:50
Sorry to hear that, it serves us very well. It's good to stick to the libraries you're the most comfortable with.
rperdon
@rperdon
Nov 29 2017 20:04
I'm still hopeful we can figure a way to get something useful trained or perhaps figure the link to get outside models loaded correctly. My nvidia support contact indicates potentially there is a problem with the weights not being mapped correctly when loaded.