lm/train.py Data Loss Error

[Muhammd-Hamza-Sabir]

I am trying to train generator model on CPU and getting error given below. Could you please provide why that error is occured:

DataLossError (see above for traceback): corrupted record at 0

[rowanz]

hi Muhammad, my guess is that you don't have the RealNews in the right format -- it only accepts .tfrecord files. However, I don't think you'll be able to train a generator model on CPU, since it's quite large.

[leehelenah]

I am trying to prepare the .tfrecord files before training the generation model. I run the prepare_lm_data.py with the generator=mega~dataset=p0.94.jsonl as input_fn.

Could you please help me to solve the following error:

Traceback (most recent call last):
File "realnews/prepare_lm_data.py", line 218, in
final_desired_size=max(args.max_seq_length + 1, 1025)):
File "realnews/prepare_lm_data.py", line 187, in buffered_and_sliding_window_article_iterator
for article in article_iterator(encoder, final_desired_size=final_desired_size):
File "realnews/prepare_lm_data.py", line 145, in article_iterator
unconditional_prob=.35)
File "/workspace/grover/sample/encoder.py", line 258, in tokenize_for_grover_training
article_pieces = _tokenize_article_pieces(encoder, item)
File "/workspace/grover/sample/encoder.py", line 175, in _tokenize_article_pieces
'article': [encoder.begin_article] + encoder.encode(item['text']) + [encoder.end_article],
KeyError: 'text'

Thanks.

[Muhammd-Hamza-Sabir]

I am getting following error while training even i have converted input files into .tfrecord. I didn't find any help regrding this error. Will you please provide what this error is about and how it could be resolved?
"absl.flags._exceptions.IllegalFlagValueError: flag --input_file=None: Flag --input_file must have a value other than None."

Thanks

[rowanz]

@helenalee1994: unfortunately that file (for discrimination) can't be used for generation. Among other things, it has additional entries, and the metadata fields have different names 😢 so for instance, I used 'text' instead of article.

would it help if I publicly shared a small subset of RealNews for debugging? for the full version, please fill out the google form on the github repo.

@Muhammd-Hamza-Sabir You need to pass in --input_file=your_tf_record.tfrecord or something like that 😄

[Muhammd-Hamza-Sabir]

@rowanz Thanks for your response.
I am doing the same but still getting these errors.
Yeah, it would be very helpful if you share a small subset of news.
Thanks

[leehelenah]

@rowanz Thanks for your quick reply. It would be a great help if you publicly share a small subset of RealNews. 😄😄😄

[kino188]

@Muhammd-Hamza-Sabir Hallo, have u solved the input_file error when converting input files into .tfrecord?