06:28 < tham> Writing a finetune class to find tune stack autoencoder

06:29 < tham> When I want to reuse the StackAutoencoderFunction and SoftmaxFunction, I find out two problems

06:30 < tham> First one is the input of StackAutoencoderFunction and SoftmaxFunction need to be changed when you finetune the parameters

06:31 < tham> second is the SoftmaxFunction need to cache the probabilities value, else you have to recalculate the probabilites two more times

06:36 < tham> What would you do?

06:36 < tham> I would like to contribute this class back to the community(if it could work), but the current design of SoftmaxFunction and StackAutoencoderFunction

06:37 < tham> cannot finish the fine tune algorithms without solving the first obstacle(second one will affect the performance)

06:39 < tham> Do you think that allow the users to change the underlying data of StackAutoencoderFunction and SoftmaxFunction is a good idea?

06:39 < tham> change the arma::mat const &data to arma::mat *data

06:40 < tham> About the second one, when you call the Gradient function, the probabilities should be calculated already, you do not need to recalculate it again(if I did not make any mistake)

06:42 < tham> if you cache it.When you finetune the parameters, you need to use this probabilities matrix one more time, without cacheing, you will need to calculate it three times

06:42 < tham> Please give me some suggestion about this one, thanks

07:08 < tham> There are another solution

07:08 < tham> Pass the reference of the input to the fine tune class

07:10 < tham> This discussion looks bizarre without codes, I think I better open a new ticket on github