reset password
Author Message
msargent
Posts: 519
Posted 12:13 Mar 03, 2015 |

For that part of the assignment (e.c.), I'm only asking for training accuracy: you don't need to divide the data into a training/test set. I got 98% accuracy: you should get around that (in the 90s). 

lcampo14
Posts: 14
Posted 14:55 Mar 06, 2015 |

So once we get our thetas which is our results then just apply forward propagation using those thetas and calculate h(x) for every x. The compare h(x) to y?

Also our thetas are dependent on the number of iterations on opt.fmin_cg function right? How long is our program suppose to run for(seconds or minutes)? When i use the default maxIteraions to 50 it takes a couple of minutes for the program to run? In the pdf it suggests to use iterations to 400 but that seems to increase the programs runtime by a longtime.

 

Would it be wrong to round the values in h(x) so that they either are 0 or 1 in order to compare them to y?

Thanks

msargent
Posts: 519
Posted 15:08 Mar 06, 2015 |

400 will take a long time to run. You should get around 95% training accuracy with 50 iterations.

Yes, just use the thetas from opt.fmin_cg and run forward propogation.

One way to do it is convert the result to an actual digit, then compare with the digit in the original Y vector (which is a vector of digits).

Last edited by msargent at 15:09 Mar 06, 2015.
lcampo14
Posts: 14
Posted 16:18 Mar 06, 2015 |

Thanks I am getting 93-95% with 50 iterations. It is taking around 5-6minutes for the program to run is that expected?

msargent
Posts: 519
Posted 16:21 Mar 06, 2015 |

That's about right.