I recently re-visited an old project of mine that uses seq2seq and recurrentshop. After updating both to the newest versions (previous last commits for each had been ~December 2016), and adjusting my code accordingly (only high-level API calls are different, no changes to core models), I find that the model no longer trains. Downgrading Keras back to 1.2 and downgrading seq2seq to commit github.com/farizrahman4u/seq2seq@223fbfdda183c2dddabb69ad0ad74bf9def69981 and recurrentshop to github.com/farizrahman4u/recurrentshop@ea67f41600b51bc8719286ae88a568be1e85956a
fixes these issues. Apparently, somewhere along the line, the models' core functionalities were broken.
I tried this using AttentionSeq2Seq, but I believe the Seq2Seq model is also broken.
I recently re-visited an old project of mine that uses seq2seq and recurrentshop. After updating both to the newest versions (previous last commits for each had been ~December 2016), and adjusting my code accordingly (only high-level API calls are different, no changes to core models), I find that the model no longer trains. Downgrading Keras back to 1.2 and downgrading seq2seq to commit github.com/farizrahman4u/seq2seq@223fbfdda183c2dddabb69ad0ad74bf9def69981 and recurrentshop to github.com/farizrahman4u/recurrentshop@ea67f41600b51bc8719286ae88a568be1e85956a fixes these issues. Apparently, somewhere along the line, the models' core functionalities were broken.
I tried this using AttentionSeq2Seq, but I believe the Seq2Seq model is also broken.