How do I fill in the loss of a hyperparam causing divergence?

[cuihenggang]

Hello,

I'm using spearmint-lite to search for a good step size and regularization term for my machine learning application. If I understand correctly, I should fill in the training loss (if I'm not interested in the validation loss) in the result.dat file after running with each set of step size and regularization term.

However, I find sometimes my application diverges (with nan or inf loss) with the step size and regularization term proposed by Spearmint, and Spearmint does not accept nan or inf values as the result. So in that case, how do I tell Spearmint that this set of step size and regularization term leads to divergence?

Thank you!
Cui

[JasperSnoek]

You should mark those as "constraint violations" (take a look at this paper
for details: http://arxiv.org/abs/1403.5607). This is implemented in
GPConstrainedEIChooser.py
https://github.com/JasperSnoek/spearmint/blob/master/spearmint/spearmint/chooser/GPConstrainedEIChooser.py

Hope that helps!

On Wed, Jul 20, 2016 at 4:10 PM, Henggang Cui [email protected]
wrote:

Hello,

I'm using spearmint-lite to search for a good step size and regularization
term for my machine learning application. If I understand correctly, I
should fill in the training loss (if I'm not interested in the validation
loss) in the result.dat file after running with each set of step size and
regularization term.

However, I find sometimes my application diverges (with nan or inf loss)
with the step size and regularization term proposed by Spearmint, and
Spearmint does not accept nan or inf values as the result. So in that case,
how do I tell Spearmint that this set of step size and regularization term
leads to divergence?

Thank you!
Cui

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#57, or mute the thread
https://github.com/notifications/unsubscribe-auth/AC7L286k2KfILz7omndELpRQF5hMPUN2ks5qXoCvgaJpZM4JRKKl
.

[cuihenggang]

Hi,

Thank you for pointing out the constrained chooser. I am now using the GPConstrainedEIChooser. However, I get the following exception. What could be wrong?

Traceback (most recent call last):
File "spearmint/spearmint-lite/spearmint-lite.py", line 219, in
main()
File "spearmint/spearmint-lite/spearmint-lite.py", line 84, in main
main_controller(options, args)
File "spearmint/spearmint-lite/spearmint-lite.py", line 190, in main_controller
np.nonzero(grid_idx == 0)[0])
File "/usr/local/lib/python2.7/dist-packages/chooser/GPConstrainedEIChooser.py", line 240, in next
durations[complete])
File "/usr/local/lib/python2.7/dist-packages/chooser/GPConstrainedEIChooser.py", line 161, in _real_init
self.constraint_gain = state['constraint_gain']
KeyError: 'constraint_gain'

Thank you very much!

Cui

[cuihenggang]

I find the 'constraint_gain' is not dumped to the state file:

spearmint/spearmint/spearmint/chooser/GPConstrainedEIChooser.py

Line 106 in b37a541

cPickle.dump({ 'dims' : self.D,

Thank you,
Cui