Make sure that stepwise prediction self-feeds one-hot encoded vectors

Klebert-Engineering / deep-spell-9

Neural Spellcheck, Autocomplete and Fuzzy-match for SQLite FTS5 🤖

MIT License

2 stars 0 forks source link

Make sure that stepwise prediction self-feeds one-hot encoded vectors #4

Closed josephbirkner closed 7 years ago

josephbirkner commented 7 years ago

Currently, the DSLstmPredictor._stepwise_predictor() is feeding itself the previous softmax probability distribution. This leads to a rapid decline in confidence for the predictions, since the network is only trained with one-hot encoded vectors.

josephbirkner commented 7 years ago

Fixed with 51aa078723e91e6c9e2840927e3eaccdcbb616aa