jcjohnson / torch-rnn

Efficient, reusable RNNs and LSTMs for torch
MIT License
2.49k stars 508 forks source link

Feature ideas: scroing, erasures, binary data #136

Open vi opened 8 years ago

vi commented 8 years ago
  1. Instead of sampling, scoring input texts. Like echo "The quick brown fox jumps over the lazy dog" | th score.lua ... -> 0.95; echo "Mq98Mcl mdsa oaqIMQosdf" | th score.lua ... -> 0.04
  2. Filling in esasures in input text. Like th fill.lua ... -text "The q___k br___ fox jump_ _____ the lazy d___" -> The quick brown fox jumps under the lazy duck
  3. Binary data handling. Workaround: convert to hex
  4. Compression. Like cat input.txt | th compress.lua ... | th decompress.lua
ChrisCummins commented 8 years ago

68 maybe relevant for your 1st request

vi commented 8 years ago
  1. Interactive mode. It shows N most probable next tokens, you choose (or override and input less probable token), iteratively.
dgcrouse commented 7 years ago

@vi, A lot of these are interesting, if you would like to implement them yourself (or if someone else wants to), please submit a pull request.