search

jakeret / tf_unet

Generic U-Net Tensorflow implementation for image segmentation
GNU General Public License v3.0
1.9k stars 748 forks source link

increase training_iters in demo toy problem #193

Closed samsammurphy closed 6 years ago

samsammurphy commented 6 years ago

A friend and I found that 20 iterations wasn't enough to converge on a solution (#192). Increasing to 50 iterations successfully executed on my old laptop and resulted in decent prediction accuracy.

We suggest increasing the iterations in the toy problem because it is very important that this example works out of the box.

jakeret commented 6 years ago

thx for the PR I also noticed that sometimes it takes a bit longer to converge that usual. Therefore I just committed a new version of the scripts/launcher.py where I set a random seed (np.random.seed(98765) ) and increased the number of iterations to 32.

If you could adapt your change to the same values I'm happy to merge

samsammurphy commented 6 years ago

@jakeret I changed training iterations to 32 and defined a set seed for the random number generator.

jakeret commented 6 years ago

LGTM