kaldi finetune model (help please)

[Monia Sayed]

Hi all,

It seems that the finetune is not very simple (according to me :-(  ). I 'am trying to use the code recommended in the issue #33: egs/aishell2/s5/local/nnet3/tuning/finetune_tdnn_1a.sh

I am following this steps.

1- At first i have prepared data using the scripts of mini_librispeech receips. I have mainly used prepare_dict.sh,  prepare_lang, align_si.sh (5 jobs)

2- I have created a folder under the  name of "exp/nnet3/tdnn_sp" and i have put the final.mdl of the model that i would to fine tune.

3- As i have computed the alignement at first, i have put them in exp/train_clean_5_ali (my finetune data is called train_clean_5). So my folder contains ali.x.gz, num_jobs, fsts.x.gz, final.mdl ( I think that this final.mdl is different from that of my model as i have obtained it after puting the model that i will fintune)

Finaly, i have faced the below  problem ( i  have put attached the log file)

-------------------- The error ------------------------

run.pl: job failed, log is in exp/nnet3/tdnn_sp_train_clean_5/log/init_mdl.log

Traceback (most recent call last):

  File "steps/nnet3/train_dnn.py", line 459, in main

    train(args, run_opts)

  File "steps/nnet3/train_dnn.py", line 297, in train

    input_model=args.input_model)

  File "steps/libs/nnet3/train/frame_level_objf/acoustic_model.py", line 88, in prepare_initial_acoustic_model

    else '{0}/0.raw'.format(dir))))

  File "steps/libs/common.py", line 129, in execute_command

    p.returncode, command))

Exception: Command exited with status 1: run.pl --mem 4G exp/nnet3/tdnn_sp_train_clean_5/log/init_mdl.log                 nnet3-am-init exp/train_clean_5_ali/final.mdl exp/nnet3/tdnn_sp_train_clean_5/input.raw - \|                 nnet3-am-train-transitions -                 "ark:gunzip -c exp/train_clean_5_ali/ali.*.gz|" exp/nnet3/tdnn_sp_train_clean_5/0.mdl

-----------------------------------------------

Hope that somone can help me ? 

[Monia Sayed]

Can someone help me please ?

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
You received this message because you are subscribed to a topic in the Google Groups "kaldi-help" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/kaldi-help/rpNd-LuMyIc/unsubscribe.
To unsubscribe from this group and all its topics, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/e46c5fbd-19a5-4da3-90db-0fc124816006n%40googlegroups.com.

[nshm...@gmail.com]

The log says the model you used for alignment is not compatible with the final.mdl you want to fine-tune.

If you try to tune some external model you'd better try the original model from the recipe first.