chiragnagpal/deep_cox_mixtures

Error: 'ValueError: probabilities contain NaN'

Closed this issue · 8 comments

flippercy commented

Hi @chiragnagpal:

A couple of extra questions on DCM:

  1. What is the definition of h and how to determine the best value of k and h?
  2. I got an error with my internal data:

k = 3
h = 100
from dcm import dcm_tf as dcm
model = dcm.DeepCoxMixture(k, h)

model, losses = dcm.train(model,
data_dev_X, data_dev_t, data_dev_y,
data_val_X, data_val_t, data_val_y,
epochs=10, lr=0.001, bs=128,
use_posteriors=False,
random_state=19850606,
return_losses=True,
patience=3)

Error: 'ValueError: probabilities contain NaN'

It happened at about the 10th epoch. How to fix it, please?

Thanks a lot!

chiragnagpal commented

would you be more comfortable with a pytorch implementation, I have one which is more stable than the current tensorflow one which I can share with you.

flippercy commented

Yes I do. Could you share it if stability is the root cause of the issue?

In addition, would you like to give me some insights on my question on k and h, please?

Best,

flippercy commented

Hi @chiragnagpal:

Could you share the new version of DCM with me, please? Is the size of the library small enough to be attached to an email?

Thank you.

chiragnagpal commented

Hi, would you be able to send me an email at chiragn@cs.cmu.edu?

flippercy commented

@chiragnagpal: an email has been sent yesterday.

Thank you.

flippercy commented

Hi @chiragnagpal:

Any updates, please?

Thank you.

flippercy commented

Hi @chiragnagpal:

Any updates, please?

Thank you.