Error in inference.ipynb, when loading pre-trained model

[hzahera]

Hi,

I am trying to use a pre-trained model from the model.zip, namely, kp20k-meng17-verbatim_append-rnn-BS64-LR0.05-Layer1-Dim150-Emb100-Dropout0.0-Copytrue-Reusetrue-Covtrue-PEfalse-Contboth-IF1_step_50000.pt

I got an error when running the inference notebook example https://github.com/memray/OpenNMT-kpg-release/blob/master/notebook/inference.ipynb

one2seq_ckpt_path = '../models/kp20k-meng17-verbatim_append-rnn-BS64-LR0.05-Layer1-Dim150-Emb100-Dropout0.0-Copytrue-Reusetrue-Covtrue-PEfalse-Contboth-IF1_step_50000.pt'

opt, unknown = parser.parse_known_args('-config %s' % (config_path))

#opt = parser.parse_args('-config %s' % (config_path))
setattr(opt, 'models', [one2seq_ckpt_path])

translator = translator.build_translator(opt, report_score=False)

Got error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_17673/2277166719.py in <module>
      6 setattr(opt, 'models', [one2seq_ckpt_path])
      7 
----> 8 translator = translator.build_translator(opt, report_score=False)

/data/OpenNMT-kpg-release/onmt/translate/translator.py in build_translator(opt, report_score, logger, out_file)
     42         else onmt.model_builder.load_test_model
     43     )
---> 44     fields, model, model_opt = load_test_model(opt)
     45 
     46     # (deprecated after dynamic data loading) added by @memray, ignore alignment field during testing for keyphrase task

/data/OpenNMT-kpg-release/onmt/model_builder.py in load_test_model(opt, model_path)
     90                             map_location=lambda storage, loc: storage)
     91 
---> 92     if opt.fairseq_model:
     93         # load a Fairseq-trained model, such as BART
     94         tokenizer = None

AttributeError: 'Namespace' object has no attribute 'fairseq_model'

Can you please help me to run a pretrained model to generate keyphrases from text?

[haseeb33]

@hzahera I experienced the same issue. That's what I did to resolve the issue and it worked. In file: onmt.model_builder In function load_test_model at line 92 modified -> if hasattr(opt, 'fairseq_model'): original -> if opt.fairseq_model:

[hzahera]

Thanks a lot for your reply. this modification works, but I got a new error

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_15784/2277166719.py in <module>
      6 setattr(opt, 'models', [one2seq_ckpt_path])
      7 
----> 8 translator = translator.build_translator(opt, report_score=False)

/data/OpenNMT-kpg-release/onmt/translate/translator.py in build_translator(opt, report_score, logger, out_file)
     63         )
     64     else:
---> 65         translator = Translator.from_opt(
     66             model,
     67             fields,

/data/OpenNMT-kpg-release/onmt/translate/translator.py in from_opt(cls, model, fields, opt, model_opt, global_scorer, out_file, report_align, report_score, logger)
    330             logger=logger,
    331             seed=opt.seed,
--> 332             kp_concat_type=opt.kp_concat_type,
    333             model_kp_concat_type=model_opt.kp_concat_type,
    334             beam_terminate=opt.beam_terminate,

AttributeError: 'Namespace' object has no attribute 'kp_concat_type'

[smolPixel]

@hzahera Going through the same struggles as you are rn. What I did for the kp_concat_type is directly modify in translator.py the line
kp_concat_type=opt.kp_concat_type, with
kp_concat_type='one2one',

Not the most elegant solution, but probably the easiest one (you can also replace with one2many if that's what you're using)

[hzahera]

Thank you for your reply. This helps me to fix the bug!

However, I got a new related error :D @smolPixel @haseeb33

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/tmp/ipykernel_17266/3398033325.py in <module>
      2     src=[text_to_extract],
      3     tgt=None,
----> 4     src_dir=opt.src_dir,
      5     batch_size=opt.batch_size,
      6     attn_debug=opt.attn_debug,

AttributeError: 'Namespace' object has no attribute 'src_dir'

[haseeb33]

@hzahera If I remember correctly, I changed opt with model_opt for this series of errors and worked for me. It took me days to make it work so I don't remember each step. But one big step was changing the opt. with model_opt because model_opt is from the start of the program and this opt is somewhere created in the middle and I don't know the usage of it.

P.S. Maybe I am wrong but it worked for me.

[hzahera]

@haseeb33 Unfortunately it doesn't work. Can you please share your code?

Thanks,

[memray]

@hzahera Apologies. I don't think the new code is perfectly compatible with the previous checkpoints. The parameters should be reloadable but some configs could be buggy.

We are already working on a tutorial and will release some new checkpoints and notebooks.

[smolPixel]

@memray Sorry to bug you, do you have any update on the tutorial and new checkpoints? Thank you very much!

[memray]

Hi @smolPixel , we are developing a new package and some demos on the basis of huggingface transformer. But no doubt it will not work for previous RNN models. I will upload some new checkpoints and update the notebook a bit, hopefully by next weekend. Please stay tuned.