These are chat archives for beniz/deepdetect

20th
Nov 2017
Emmanuel Benazera
@beniz
Nov 20 2017 10:16
We've updated the Docker GPU image so that it supports Pascal architectures by default
(there was a conflict with some drivers at a point, but this should have disappeared by now)
jubeenshah
@jubeenshah
Nov 20 2017 11:38
Yeah I'm pulling the pascal image right now. Was getting a timeout error unitl yesterday night, Thanks a tonne for the support, Will let you know how it proceeds.
rperdon
@rperdon
Nov 20 2017 15:09
Just some clarification iteration versus epoch. In digits, 1 epoch is a full run through all the test material once. Digits hits some good accuracy after 25 epoch on average. What is the measure of iterations? Is it also 1 full traversal through the training image list? I have 34k images to train on and if the iterations was based on 1 image test, this would explain my current 50/50 split dilemma for the results after 30k iteration.
Emmanuel Benazera
@beniz
Nov 20 2017 15:27
You divide the number of training samples by the batch size to get the number of iterations that represent one epoch.
rperdon
@rperdon
Nov 20 2017 16:07
So of 34k, I have 30k images in training (0.1 validation), which with 16 batch size means I have only 1900~ per epoch.
So if I change my batch size to 1, then 30k per epoch?
So would I be correct at this point with 30k training images and 1 batch size that 1 iteration = 1 epoch?
rperdon
@rperdon
Nov 20 2017 16:14
So in my previous training attempts, was it only training on the same 1900 images per iteration for 30k iterations? It would explain why I was only getting a 49% accuracy rate if the training sample per epoch was that small.