You probably meant to initialize the input->hidden weight matrix with the result of unsupervised training on the full corpus. A little tweak on how these weight

You probably meant to initialize the input->hidden weight matrix with the result of unsupervised training on the full corpus. A little tweak on how these weights are initilized would do: https://github.com/facebookresearch/fastText/blob/master/src/fasttext.cc#L358 https://github.com/facebookresearch/fastText/blob/master/src... I was a bit curious why they did not offer this by default. It seems quite useful.