Can't restore model from pre-trained model link about f-lm HOT 5 CLOSED

Just wanted to follow up - I started evaluation on the full dataset and it looked like the perplexity was consistently getting below 25, so that is good. I didn't evaluate to the end because I didn't want to wait, but it looks good. Many thanks!

from f-lm.

In your "--hpconfig", num_of_groups=0, thus it is trying to restore params for LSTM-based model.
The link contains a snapshot of GLSTM based model (with num_of_groups=4) described in the paper. Thus you have to do:
--hpconfig run_profiler=False,float16_rnn=False,max_time=$SECONDS,num_steps=20,num_shards=8,num_layers=2,learning_rate=0.2,max_grad_norm=1,keep_prob=0.9,emb_size=1024,projected_size=1024,state_size=8192,num_sampled=8192,batch_size=128,num_of_groups=4
Also, make sure you use TF 1.0

from f-lm.

Thanks.

Succesfully loaded model at step=2262215.
0: 4.596 (99.066) ... 
Results at 2262215: log_perplexity = 4.596 perplexity = 99.066

I am using tf 1.4 and it seems to be working with that version too

from f-lm.

hm, TF RNN API have changed significantly from 1.0 to >=1.3. Not sure if it'll work correctly, btw, perplexity on 1bw should drop to < 25

from f-lm.

Thanks for the info. I didn't actually evaluate with the real dataset, I just wanted to make sure it compiled for me. I will try later on and if the results are >25 I will try on v 1.0.

from f-lm.