Hi,
I just downloaded this code and tried to work around with some examples. However, I got the following errors:
(I use python 3.5)
Traceback (most recent call last): File "bayes_opt.py", line 444, in <module> num_iterations=args.num_iterations) TypeError: 'module' object is not callable
Thanks for any help!

Within the jensen_shannon_divergence function in the similarity file should the following line be this:

sim = 1 - 0.5 * (scipy.stats.entropy(repr1, avg_repr) + scipy.stats.entropy(repr1, avg_repr))

and not this:

sim = 1 - 0.5 * (scipy.stats.entropy(repr1, avg_repr) + scipy.stats.entropy(repr2, avg_repr))

learn-to-select-data/bilstm_tagger/src/simplebilty.py

from bilstm_tagger.src.lib.mnnl import FFSequencePredictor, Layer, RNNSequencePredictor, BiRNNSequencePredictor
from bilstm_tagger.src.lib.mio import read_conll_file, load_embeddings_file

bilstm_tagger.src.lib is missing
In the original repo, this library exists.

Hi @sebastianruder ,
I'm currently trying to reproduce your experiment using the SANCL 2012 data. I have obtained the SANCL 2012 LDC but I am not clear how did you split the data into train, dev, and test.
In the LDC data they don't provide the split and also the CONLL format of the annotation.

Can you explain in more details about this? It would be helpful if you have the scripts that process the SANCL 2012 data that construct the split and conll data format that are used by your code.

Hi,

The article is quite fascinating and very impressing in its methodological rigor and the breadth of validation tasks. And now, coming to adapt the method for a task that is not one of the ones used for validation in the paper ― e.g. a classification task (and not sentiment analysis, part-of-speech tagging, or parsing), I wonder whether anyone may kindly point out the point in the code to intervene for this sake, or more generally, how to use this project for a scenario where the task is not one of the three employed here.

I ask this not to avoid studying the code per-se, but rather as I'm not currently sure from the project structure, whether there's a certain pipeline of executions (of the different python executables) that needs to be taken in order to materialize a result. The documentation is very good, but somewhat oriented to reproducing the specific experiments, whereas I wish to try out the code, as mentioned above, for a different downstream task...

I am eager to hopefully have your comments, regarding this great work vis-a-vis my use case for it.

Thanks in advance!

is the 'features' module in the import of bayes_opt.py missing?