Thanks for your sharing!
I have now applying your code for clothing classification, when I pretrain the network ( ft_last_layer.prototxt) using the script (ft_last_layer.sh ), and the number of training data is about 220K with ID 30K+, it can not converge. The initial loss is about 10, but after a lot of iterations about 100k+, the loss is still about 9. At the mean time , I have made another experiment, all the environment keeps the same as previous experiment except the number of training data become small(about 7K with 800ID) ,which is extracted from the previous training data, now it converge quickly.
So my question is, when the training data is large, why does it not converged? are there existing some tricks to make its convergence?
Thank you very much!
Thanks for your sharing!
I have now applying your code for clothing classification, when I pretrain the network ( ft_last_layer.prototxt) using the script (ft_last_layer.sh ), and the number of training data is about 220K with ID 30K+, it can not converge. The initial loss is about 10, but after a lot of iterations about 100k+, the loss is still about 9. At the mean time , I have made another experiment, all the environment keeps the same as previous experiment except the number of training data become small(about 7K with 800ID) ,which is extracted from the previous training data, now it converge quickly.
So my question is, when the training data is large, why does it not converged? are there existing some tricks to make its convergence?
Thank you very much!