Cool ! 100% correct for digits
StevenLOL opened this issue · comments
Reach 1 at 800 iteration, then drop to 0.92 -> 0.96 and return to 1.
May I try on words?
I am so glad someone was able to reproduce near 100% accuracy :)
On second thought it's maybe not that impressive, given the limitation of the task. but still it is probably better than those useless Linux speech recognition frameworks and thrillingly motivating!
I uploaded the training data for English words and an experimental corresponding net.
However it is work in progress and a very different beast.
Any assistance in applying recurrent networks and multi-label sliding windows is much appreciated.
On Dec 21, 2014, at 3:49 PM, Steven notifications@github.com wrote:
Reach 1 at 800 iteration, the drop to 0.92 -> 0.96 and return to 1.
May I try on words?
—
Reply to this email directly or view it on GitHub.
Indeed the accuracy is reported by caffe during training.
I was unable to reproduce the 100% on training or testing set.
My prediction, ported from "caffe/distribute/python/classify.py" always report the class 3, and thus 10% accuracy.
Any idea how the 100% accuracy come from?
I am so disappointed about caffe, the so called state or art project lack of basic function such as "predict".
classify.py doesn't handle grayscale images well that's why you can't get meaningful results yet.
I will post some code so that you can reproduce the results of training "live".
Or you take it as an exercise from the ipython notebooks. Actually it's much simpler then trying to modify the classify.py file, which is really only optimized for the imagenet example and not really usable for all the other examples.
I agree that it's pretty poor that the basic caffe app lacks the 'predict' functionality!
but I I guess it's up to us to add it if we want.=