Different results from the command line fasttext

[olegtarasov]

Different results from the command line fasttext

[olegtarasov]

Hi @sidc!

Sorry for the late reply. This library uses custom-built fastText binary, which quite closely follows the master branch of the official fastText repository. The fact that you use 0.9.1 binary and get different results from nn is weird.

I've just tried to train an unsupervised model with latest fastText bin from master (0.9.2) and then query nearest neighbours from that binary and from the .NET wrapper (which uses 0.9.1). The results are identical:

Query word? slow
low 0.999476
overcooked 0.998307
hours 0.998296
cookbook 0.998271
cookbooks 0.998237

[0] = {FastText.NetWrapper.Prediction}
 Label = {string} "low"
 Probability = {float} 0.999475658
[1] = {FastText.NetWrapper.Prediction}
 Label = {string} "overcooked"
 Probability = {float} 0.99830699
[2] = {FastText.NetWrapper.Prediction}
 Label = {string} "hours"
 Probability = {float} 0.99829638
[3] = {FastText.NetWrapper.Prediction}
 Label = {string} "cookbook"
 Probability = {float} 0.998271048
[4] = {FastText.NetWrapper.Prediction}
 Label = {string} "cookbooks"
 Probability = {float} 0.998237133

Can you describe your test case with maximum detail, so I would try to reproduce it?