why is loss_nce NaN through out the training phase ?

[vijeth8]

No description provided.

[meereeum]

@vijeth8 can you provide more context ? i don't see this.

[kkrasnas]

I have this problem for my tests. I use Bloomberg news from financial news dataset https://github.com/philipperemy/financial-news-dataset, where I am taking only 10 documents for testing. And I want 10 documents clustered into 5 topics.
Particularly when the window equals 4, loss_nce is often NaN, and for window size 5 I get an error:

model.py", line 253, in make_feed_dict
    rand = np.random.uniform(0, 1, doc_is_same.shape[0])
AttributeError: 'bool' object has no attribute 'shape'