Option to make the nets more deterministic

[run2]

What do you think about introducing an option to seed the random before doing the KFold on train test split ? That way the net predictions and loss details will be more deterministic over multiple runs on the same set.
Thanks

[dnouri]

I think right now you can do a np.random.seed(42) before training and it'll be deterministic?

[run2]

Sure yes. I just wanted to check if you want to add it as a option. I will close this and you can reopen it if and when you want to.

[cancan101]

The default for KFold is actually NOT to shuffle (i.e. it is deterministic):
https://github.com/scikit-learn/scikit-learn/blob/38104ff4e8f1c9c39a6f272eff7818b75a27da46/sklearn/cross_validation.py#L320

[dnouri]

Ah yes, very useful if your data is not independently distributed.

[cancan101]

It might be worth tossing a note in the nolearn docs along the lines of:

By default no shuffling occurs, including for the (stratified) K fold cross- validation....Keep in mind that train_test_split still returns a random split.

Related: #12

[dnouri]

Yes, you're right. And that reminds me that I should be working on proper docstrings soon.

[neilsummers]

From what I can see there is more randomness than just in the train test splits. From the tests I have done there is randomness when I include a dropout layer. I can't seem to make this the same for each run by doing a np.random.seed(42) before the run. I have tried tracing it back throug the source, and it appears to be setting a seed by default in RandomStreams from theano, which is called in the DropoutLayer, but I am still getting changes from run to run when I include a DropoutLayer, but no changes when there is no DropoutLayer.

[dnouri]

There's an issue for that in Lasagne: Lasagne/Lasagne#6