Save/load trained model

[kjancsi]

Hi, is there any way to save/load trained models? Thanks.

[jyt109]

check out pickle. https://docs.python.org/2/library/pickle.html

[kjancsi]

Thanks. I tried pickle'ing the model object in tutorial.py after it has been trained, but that did not work.

[Newmu]

Looking into this, on the same device (cpu or gpu) pickle/joblib should work but almost certainly won't across devices. We are planning an explicit save interface to take care of all the edge cases to make sure this just works.

[Slater-Victoroff]

@kjancsi Just wanted to give a status update from the sidelines. Looks like Alec has figured out what the issue is, and is just working on a couple slightly painful implementation details here.

@Newmu Something this week?

[Newmu]

Load and save functions are now added to passage.utils. Example usage in readme.

Let me know if you have any issues - supports cross device, so you can train on a GPU and predict on a CPU and vice-versa.

[kjancsi]

Thanks for implementing this feature. I've noticed an issue when using OneHot input representation.

Traceback (most recent call last):
  File "....py", line 98, in <module>
    model = load('save_test.pkl') # How to load
  File "/.../Passage/passage/utils.py", line 44, in load
    model['config']['layers'] = [getattr(layers, layer['layer'])(**layer['config']) for layer in model['config']['layers']]
TypeError: __init__() got an unexpected keyword argument 'weights'

[Newmu]

Thanks! Weights is used to set parameters when loading a model and forgot the onehot layer doesn't have any! Quick fix pushed.

[kjancsi]

Thanks, it works now.