see-- / natural-question-answering Goto Github PK

When I am trying to run the demo.py file on Google Colab, I am getting the following error with the tokenizer.

ValueError: Non-consecutive added token 'td_colspan' found. Should have index 30522 but has index 1 in saved vocabulary.

Please help me resolve this error.

Hi everyone,
I try to run this repo, but I met error below:

Traceback (most recent call last):  
  File "train_eval.py", line 481, in <module>
    main()
  File "train_eval.py", line 437, in main
    num_added = tokenizer.add_tokens(add_tokens, offset=offset)
TypeError: add_tokens() got an unexpected keyword argument 'offset' 

My environment:

• Python: 3.7.0
• Transformers: 2.2.0
• Tensorflow: 2.0

I checked all of version of the transformer, but I haven't found version that has offset argument in the add_tokens method

Hi,

demo is only for short answers and not long answers?

Thanks
Mahesh

First of all thank you for the great work. Your Q&A model rocks. Really interesting to see what is possible for next level Q&A.

I played around with the model in tf-hub and noticed it has a memory leak.
Here is my code:

import tensorflow as tf
import tensorflow_hub as hub
from transformers import BertTokenizer

tokenizer = BertTokenizer.from_pretrained("tokenizer_tf2_qa")
model = hub.load("https://tfhub.dev/see--/bert-uncased-tf2-qa/1")

for question, context in data:
    
    # create input vector representation
    encoded = tokenizer.encode_plus(question, context, add_special_tokens=True)
    input_word_ids = encoded["input_ids"]
    input_mask = encoded["attention_mask"]
    input_type_ids = encoded["token_type_ids"]

    # convert to tf.int32 and pass through model
    input_word_ids, input_mask, input_type_ids = map(
        lambda t: tf.expand_dims(tf.convert_to_tensor(t, dtype=tf.int32), 0),
        (input_word_ids, input_mask, input_type_ids),
    )
    outputs = model([input_word_ids, input_mask, input_type_ids])

I tested for both Tensorflow 2.1.0 and 2.2.0 on a cpu machine.

I wonder if this warning is related to the memory leak:

WARNING:tensorflow:5 out of the last 5 calls to <function recreate_function.<locals>.restored_function_body at 0x7f23a70d9680> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings is likely due to passing python objects instead of tensors. Also, tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. Please refer to https://www.tensorflow.org/tutorials/customization/performance#python_or_tensor_args and https://www.tensorflow.org/api_docs/python/tf/function for more details.

Any idea what could be the problem?

Hi see--

I am trying to add tokens by
tokenizer.add_tokens(add_tokens, offset=offset)

But I got error

TypeError: add_tokens() got an unexpected keyword argument 'offset'

Are you using anything different?

Regards,
Ankur

I'm trying to run the tfhub sample, got the following error
Model name 'tokenizer_tf2_qa' was not found in tokenizers model name list
(I did pip install transformers),
can you help?

Hello,
I am getting following error while executing demo.py with custom document text(more text) on CPU. It works fine with GPU though.

tensorflow.python.framework.errors_impl.InvalidArgumentError: indices[0,512] = 512 is not in [0, 512)
[[{{node StatefulPartitionedCall/StatefulPartitionedCall/tf_bert_for_natural_question_answering/StatefulPartitionedCall/bert/StatefulPartitionedCall/embeddings/position_embeddings/embedding_lookup}}]] [Op:__inference_restored_function_body_89164]

Thanks
Mahesh

Is there a pretrained model available?