Stanislaw Jastrzebski

Another fix: baseline dist calculation should use float as: closest = np.abs(last_q - last_a) / (float(len(passage)) - 1). I understand stopwords removal is done during data preprocessing

Here is how to "fix" it without changing Blocks code https://gist.github.com/kudkudak/f8d80e1113afbc34b8aa5efac498ce82. 1. https://gist.github.com/kudkudak/f8d80e1113afbc34b8aa5efac498ce82#file-gistfile1-txt-L83 adds more parameters to save to Checkpoint 2. https://gist.github.com/kudkudak/f8d80e1113afbc34b8aa5efac498ce82#file-gistfile1-txt-L101 makes sure they get deserialized Alternatively one could...

Sketch of simplest version can be found here https://gist.github.com/kudkudak/e76bce480b07c4ed0e4ebb39f3cecd93

https://github.com/lucy3/grounding-embeddings

Have you tried normalizing the word vectors before passing them to the analogy solver? This could help

I am not sure I fully understand. Is your suggestion that we support gensim Embedding class?

We are thinking about a release on a short timescale, we will probably indeed change the name :) Thanks for the suggestion.

Actually noticed the same thing here https://github.com/facebookresearch/mixup-cifar10/issues/3 . On top of that it seems to me that the final BN layer is missing.

@cieplinski-tobiasz Could you take a look?

👍 , also please allow for name with subdirectories, for instance I would run sth like ``` smart_dispatch.py -n 2016_11_1/my_experiment_with_batch_norm ``` , as for instance I like to group experiments...