Conversation
python/config.cfg
Outdated
| # cost function: square_loss, binomial_cross_entropy, softmax_loss, auto | ||
| # auto mode will match the out_type: boundary-softmax_loss, affinity-binomial_cross_entropy | ||
| cost_fn = auto | ||
| cost_fn = square_loss |
There was a problem hiding this comment.
@nicholasturner1 why do you change it to square loss? I think that auto is better for default setting.
There was a problem hiding this comment.
The original thought was that it would be easier for new people to interpret/troubleshoot square error instead of having to take an extra step to figure out which error they're even using. Not a huge deal for me either way though. If you greatly prefer auto I'll swap it back
There was a problem hiding this comment.
yes, I prefer auto. square loss corresponds to linear output. To help the new users, we should explain these in the comments in config.cfg file though.
|
@nicholasturner1 could you solve the conflict? current code did not pass the autotest in travis. |
|
Nope, I hadn't changed things yet still (sorry...), but I also didn't merge it. @sachinravi14, do you know what the issues were? |
|
@sachinravi14 did you make any fix of merging this pull request? |
|
No, I just merged and it worked for both synapse and semantic branches. Also, I think masking also needs to be fixed on test because I don't think the voxels are being counted correctly there. |
3 participants
Addressing #66 : Consistently applying mask across error computation, cls error, and testing error/cls
Two further commits: