Torch/nn update breaks loading OpenNMT models trained with earlier version(?)

Support
1

Having recently updated our torch installation and packages we are no longer able to load the models that were trained using the old version:

$ translate.lua -model model_old.t7 …
…
[05/24/17 08:50:38 INFO] Loading ‘model_old.t7’…
[05/24/17 08:50:47 INFO] Model seq2seq trained on bitext
…/torch/install/bin/lua: …e/local/src/torch/install/share/lua/5.2/nn/JoinTable.lua:38: bad argument #1 to ‘copy’ (sizes do not match at /tmp/luarocks_cutorch-scm-1-5403/cutorch/lib/THC/THCTensorCopy.cu:31)

Models trained in current setup work fine. Is it expected behaviour or does anoyne have some similar experience?

2

Hello, we are trying to be careful with backward compatibility. Do you know with which version of OpenNMT you trained the “old models”?

3

Related to:

On which version did you train model_old.t7 and with what options?

4

I used the same OpenNMT version for the “old” models as I’m using now, in which the last commit is 6521cc0cd4d64d07bb4d94ca2b99787f1cd66b7d, Wed Apr 19 12:23:11. It’s only the torch (and its packages) installation that I updated, with ‘old version’ I was referring (pretty misleadingly, sorry for that) to the torch stuff.

Training options (it’s just a toy model):

-src_vocab_size 50000 -tgt_vocab_size 50000 -src_seq_length 80 -tgt_seq_length 80
-brnn -layers 2 -rnn_size 500 -rnn_type LSTM -word_vec_size 500 -feat_vec_exponent 0.8 -max_batch_size 32 -input_feed 1
-end_epoch 18 -optim sgd -learning_rate 1 -dropout 0.3 -learning_rate_decay 0.5 -start_decay_at 9 -curriculum 0

5

See https://github.com/OpenNMT/OpenNMT/issues/284#issuecomment-303749651

6

thanks for your great support!
now, just need to be upgraded nn,It worked fine.
luarocks install nn

7

Thanks, all is back to normal now.