This is my own data as on hackster where my previous model had 92% accuracy
Trainng on floydhub GPU
Helper to extract class label
Error rate down to 3.4%
Lets save the state of the model
It appears I have a couple of incorrectly labelled images, not too surprising as I labelled them manually. Occasionally it is confusing rain for birds..?
Lets train the whole model and see if accuracy improves..? Probably I should clean the dataset first but I am impatient
Error rate up, just as in the tutorial. Lets find the best error rate
OK the learning rates in range 1e-6,1e-3 look like a good bet
That's an improvement already!
Not much in it. I wont waste compute on this dataset, but anyways we are doing better than classificationbox :)