the traing loss won't decrease for LeviHassner

[haoyu09]

Hi, I train just as README $ python train.py --train_dir /home/dpressel/dev/work/AgeGenderDeepLearning/Folds/tf/gen_test_fold_is_0 --max_steps 10000 But the loss doesn't decrease at all. Is there something wrong with the method?

[dpressel]

The documentation on the page for running showed 10k steps on gender. This is not necessarily the recommended number of steps, and the learning rate was not specified in the example so it defaulted to the program defaults.

Since I wrote the original docs, I increased the learning rate default, and added a learning rate decay schedule, and set the number of default steps higher (these params worked well for age). I have updated the docs to explicitly lower it back like this --eta 0.001 for gender training (as it was before) and you should see the loss decreasing.

And you can try running with the defaults by removing the --max_steps or increasing it. The learning rate will then follow a staircase decay schedule.

The program allows you to control all of the hyperparameters.

[dpressel]

I have updated the docs to hopefully minimize confusion in the future