adbrebs/spynet

Neural networks library on top of theano

spynet's Issues

Network storage

Consider storing a network's hyperparameters in a JSON or YAML file instead of hdf5.
hdf5 could be only used for storing weights.

Monitor vectors

So far, each Monitor can only monitor a single scalar. Would be nice to extend to vectors

trainer.py

In trainer.py, you define: self.n_train_batches = ds.n_train / self.batch_size
What happens if ds.n_train is not a multiple of self.batch_size? Eg n_train=166 and batch_size=20, then n_train_batches=8. So we are not using the last 6 elements in the training set.