make-skip-thought-model
dumps.encoder
as well as whole skip-thought modelskip-thought-vector
returns actual word sequence- rare word will be converted to special symbol
<unk>
- rare word will be converted to special symbol
- function
make-skip-thought
->make-skip-thought-model
- variable
shareed?
->shared-decoder?
- add optimizer
- move
clip!
to optimizer clip!
limits gradients by using l2-normrewrite!
->update-param!
- add clip! to encoder-decoder models to avoid NaN
- changed clipping value from -1 to 1
- skip-thought model skips incomplete conversation line
- trivial changes and fixes
- update documentation
- dependencies
[org.clojure/clojure "1.9.0"]
[net.mikera/core.matrix "0.62.0"]
[clj-time "0.14.2"]
[org.clojure/core.async "0.4.474"]
- binary-classification-error
- throws exception wheninvalid expectation given
- valid-embedding?
- merge-param!
- attention encoder-decoder
- GRU
- LSTM
- standard rnn
- Gated Recurrent Unit(GRU)
- encoder-decoder with GRU
- Skip-Thought
- LSTM was moved to src/prism/nn/rnn/
- encoder-decoder-lstm was moved to src/prism/nn/encoder-decoder/
- sequential-output was renamed to forward (lstm)
- encoder-decoder model formally (LSTM only)
- tentatively add batch normalization of feedforward
- trivial changed architecture of each models
- use clojure.core.matrix.random for sampling
- matrix to be more readable
- error-function for multiclass classification
- add LSTM demo for multiclass classification
- add CHANGELOG.md (this file)
- now prism uses clojure.core.matrix as default matrix implementation
- add orthogonal-initialization for recurrent connection