Closed sameerkhurana10 closed 6 years ago
It seems that the model is corrupted. Looks like the HDF5 library throws a KeyError when trying to read the vocabulary from the model. So the problem is in training, not scoring. Is there something suspicious in the train log?
probably right. Other models are fine. I got bus error for this model. I think nothing to do with TheanoLM.
hi,
i am getting the following error while calculating ppl on the test set:
score command:
srun -p gpu --gres=gpu:1 theanolm score exp/blstm256_voc80k_blstm/nnlm.h5 data/rnnlm_data_all/test.dat --output perplexity --log-level debug
train command:
theanolm train exp/blstm256_voc80k_blstm/nnlm.h5 --training-set data/rnnlm_data_all/transcript.dat --vocabulary data/rnnlm_data_all/input_80000.vocab --vocabulary-format words --sequence-length 25 --batch-size 32 --optimization-method adagrad --stopping-criterion no-improvement --cost cross-entropy --learning-rate 1 --gradient-decay-rate 0.9 --numerical-stability-term 1e-6 --num-noise-samples 1 --noise-distribution unigram --noise-dampening 0.5 --validation-frequency 1 --patience 0 --min-epochs 1 --max-epochs 15 --random-seed 1 --log-level debug --log-interval 200 --gradient-normalization 5 --architecture ../architectures/word-blstm256.arch --validation-file data/rnnlm_data_all/dev.dat
just checking the size of the model:
161k
. looks suspiciously small.What does this error mean?